Skip to content

Sheet 00 / 08 · Cover · Rev 2026-10-19

Solaimuthu A(Solai)

I build multi-agent systems that show their work.

Software Development Engineer II at Blackstraw.ai, Chennai. Promoted from SDE I in Jul 2026.

Evidence
Why-Graph: agents that must prove why production broke as a graph path. Blocked a repeat at safe_min 453 ms · 167 tests · with Shoban R. · Oct 2026 · judging pending. Try to break it →
Budget
RAG in Goa: retrieval P95 76.85 ms live over 215,608 chunks, inside a 200 ms deadline · team of 3 · Aug 2026. Race it →
Award
First Place, ETH San Francisco (Coinbase Developer Platform Track) for Urbexium. Devfolio →
Drawn by
Solaimuthu A · B.E. CSE, LICET · CGPA 9.04 / 10

graphite = what happened · green = the plan, drag it · vermilion = my call, checked

Pencil drawing of a small bird landing on a branch with both wings fanned wide, construction grid faintly visible.
Wings spread · pencil · S.A.

Index of sheets

  1. 01Look CloselyA village near Kallakurichi · before 2022Pencil came before code. Drawing taught me to look for construction lines first.
  2. 02Stayed BackChennai · LICET · 2022–2026I stayed back after class: Pattarai, FabLab, go-karts, first hackathon wins.
  3. 03WingsToulouse, France · Jun–Jul 2025A student-team study re-engineering a regional airliner around hydrogen. My part: the fuel cells.
  4. 04ProductionBlackstraw.ai, Chennai · Aug 2025–nowProduction multi-agent systems: a supervisor, specialists, scoped tools, and a human before every write.
  5. 05EvidenceGraph Hacks 2026 · Oct 2026Why-Graph: agents must prove why production broke as a graph path, then block the repeat.
  6. 06BudgetHacker House Goa 2026 · Aug 2026RAG in Goa: retrieval inside a hard 200 ms deadline. We measured a reranker and cut it.
  7. 07Loose PagesEverywhere · 2024–2026Just start: a dated ledger of hackathon builds, and ProjectX, the one I built for myself.
  8. 08Your TurnHere · nowAsk the sketchbook anything about my work, or leave a note.

Sheet 01 / 08 · A village near Kallakurichi · before 2022

Look Closely

In short: Pencil came before code. Drawing taught me to look for construction lines first.

Place
A village near Kallakurichi, Tamil Nadu
Medium
Graphite pencil, signed in red ink
Habit
Construction lines first
Sheet
01 / 08

I'm from a small village near Kallakurichi. I draw in pencil (animals mostly) and I sign every sheet in red ink.

Drawing gave me the one habit I still use every day. You don't start with the feathers. You start with the construction lines: the circles and axes nobody sees in the finished drawing, but which decide whether it stands up at all. A system is the same. Get the structure right first and the details have somewhere to sit.

That's why this site is a sketchbook. Each sheet starts in pencil, then turns into something you can actually use.

Heavily shaded pencil close-up of an owl's eye and half its face, with layered feathers and a hooked beak.

Look closely1/6

Sheet 02 / 08 · Chennai · LICET · 2022–2026

Stayed Back

In short: I stayed back after class: Pattarai, FabLab, go-karts, first hackathon wins.

Place
LICET, Chennai
Degree
B.E. Computer Science and Engineering, 2022–2026
CGPA
9.04 / 10
Community
LICET Pattarai · LICET FabLab · Team Turbonites
Internship
Skill First Labs, full-stack (MERN), Jul–Sep 2024
Award
First Place, ETH San Francisco (Coinbase Developer Platform Track), Urbexium, $2,500
Sheet
02 / 08

Coming to Chennai for college was a big step. Everything was new, and honestly I had a lot of self-doubt.

One small decision changed it. Instead of going back to my room after class, I started staying back: watching, learning, trying. That's how I ended up in LICET Pattarai, the student tech community, where I later ran hands-on AI and prompt-engineering workshops for 200+(self-reported) students; at the LICET FabLab, helping students use the 3D printers and laser cutters; and with Team Turbonites, building go-karts from scratch, where I learned machine tools and CAD.

Projects led to workshops, workshops to hackathons. My biggest win so far was Urbexium, a public grievance platform that took First Place at ETH San Francisco on the Coinbase Developer Platform track. Along the way there was CivicFix with Merlyn, Best Subject Matter Expert at Design Spark 2025, and runner-up finishes at the TechnoTeen hackathon and the Vmaker AI product teardown.

The moment that mattered most wasn't a prize, though. It was bringing my parents from the village to Chennai and watching them in the audience when I received the LICET Star Award.

Every build from these years is in the ledger on Sheet 07.

Pencil drawing of two clasped hands forming an anatomical heart, with a rose blooming from the top.
Solaimuthu presenting on stage at the Vmaker AI product teardown, a slide behind him.
Vmaker AI product teardown, runner-upreal photo, not generated

Sheet 03 / 08 · Toulouse, France · Jun–Jul 2025

Wings

In short: A student-team study re-engineering a regional airliner around hydrogen. My part: the fuel cells.

Programme
ICAM summer programme, Toulouse (Jun–Jul 2025)
Team
Team McQuade: Dennyson, Gopika Sree, Pavithra, Solaimuthu
My part
Hydrogen: fuel-cell stack, hydrogen storage, safety systems
Status
Student-team design study, not certified
Sheet
03 / 08

In the summer of 2025 I spent four weeks in Toulouse on the ICAM summer programme. Our four-person team, Team McQuade, took on a question that sounds simple: could a 72-seat regional turboprop, the ATR 72-600, fly on hydrogen fuel cells plus batteries without growing heavier?

My part was the hydrogen: the fuel-cell stack, hydrogen storage and the safety systems around them. Dennyson led systems integration, Gopika Sree took propulsion and Pavithra took energy systems.

It was a student design study, not an aircraft programme, and I'll show it that way. The numbers below are our team's published estimates, exactly as published. One row in that table doesn't add up, so it's marked rather than quietly fixed.

Solaimuthu smiling in front of an ornate fountain in France.
France, summer 2025real photo, not generated
Mission energy profile (student-team estimate)

Student-team design estimate · ICAM summer programme 2025

Team McQuade: Dennyson, Gopika Sree, Pavithra, Solaimuthu

Flight time
3 h 05 min
Cruise
510 km/h at 25,000 ft
Takeoff weight
22,500 kg team’s design target

Drag, use the arrow keys, or pick a phase.

2,500 kW

phase power · Takeoff

Phase 1 of 5

Takeoff

Duration
5 min
Power
2,500 kW
Energy, as published
208.3 kWh
Cumulative
208.3 kWh of 3,258.1 kWh

Check: 2,500 kW × 5 min = 208.3 kWh, matches the table to its 0.1 kWh rounding.

Cumulative energy, filling toward the published total

Through Takeoff: 208.3 kWh (+208.3 kWh this phase)

Published total 3,258.1 kWh

Team’s stated source · Takeoff

“batteries for high-power phases (takeoff)”

Team McQuade’s project site, “Key Innovation”. Quoted, not computed: this sheet does not split power between sources.

The team's architecture: H₂ storage feeds a PEM fuel-cell stack, which with a battery pack (peak power support) feeds power electronics and two electric motors. Highlighted: the battery pack, named in the team's statement for takeoff.H₂ storagePEM fuel-cellstackpower electronicsbattery packpeak power supportnamed for takeofftwo electric motors
The team’s architecture, redrawn with generic part names.

Takeoff, phase 1 of 5: 2,500 kW for 5 min, 208.3 kWh; cumulative 208.3 of 3,258.1 kWh. Team's stated source: "batteries for high-power phases (takeoff)".

Simulation using recorded numbers from the team's published mission energy table (hybrid-atr-revival)
Show the numbers as a table
Team McQuade’s published mission energy table, with cumulative energy, a power × time check and the team’s stated source
PhaseDurationPowerEnergy (published)CumulativePower × timeTeam’s stated source
Takeoff5 min2,500 kW208.3 kWh208.3 kWh208.3 kWh, matchesbatteries for high-power phases (takeoff)
Climb10 min2,000 kW333.3 kWh541.6 kWh333.3 kWh, matchesnot stated
Cruise2.5 h1,000 kW2,500 kWh3,041.6 kWh2,500 kWh, matchesfuel cells for sustained cruise operations
Descent20 min500 kW166.5 kWh3,208.1 kWh166.7 kWh, RE-CHECK (table says 166.5)not stated
Reservenot listednot listed50 kWh3,258.1 kWhnot applicablenot stated
Published total3 h 05 min3,258.1 kWh3,258.1 kWh

Read the case study →

Sheet 04 / 08 · Blackstraw.ai, Chennai · Aug 2025–now

Production

In short: Production multi-agent systems: a supervisor, specialists, scoped tools, and a human before every write.

Company
Blackstraw.ai, Chennai (on-site)
Role
SDE I (Aug 2025) → SDE II (Jul 2026)
Stack
Python · LangGraph · A2A · MCP · Postgres · SSE · GCP
Checked by
28+ tests I wrote for authorization, stale writes and pagination
Sheet
04 / 08

At Blackstraw.ai I build production multi-agent systems. I joined as SDE I and was promoted to SDE II in July 2026.

The shape of the system is one most agent teams will recognise: a supervisor that routes work to specialist agents, specialists that can act only through scoped tools, and a human who approves anything that writes. My parts are the unglamorous ones that make it hold up in production:

  • Durable state. Runs and conversations live in Postgres with optimistic concurrency. A stale write is rejected by revision, not hidden by a lock. Large lists use keyset pagination; search uses full-text indexes.
  • Two database drivers, one data layer. One checkpointing library and the cloud database connector disagreed about drivers, so I built a second async path rather than give up IAM authentication.
  • An audit trail that stores no secrets. Every tool call is logged with its latency and a SHA-256 digest of its arguments. The raw arguments are never stored.
  • One way to stop. A user pressing Cancel and a browser dropping the stream both end in the same shutdown path, so nothing is left half-running.
  • A second front end, zero backend changes. I wrote a protocol adapter so a different UI can drive the same agents.

I can't show you the real system, so below is a reconstruction of the patterns, built from scratch.

Production multi-agent patternA client streams to a supervisor, which routes to three specialist agents. Specialists act only through scoped MCP tools. Any write goes through human approval before reaching durable state. Every tool call is recorded in an audit trail with its latency and a digest of its arguments. Cancelling and a dropped stream share one shutdown path.Any clientUI · SSE streamSupervisorroutes the workSpecialist ASpecialist BSpecialist CScoped MCP toolsread · propose-writeHuman approvalbefore any writeDurable staterevisions · keyset · FTSAudit traillatency · sha256(args)One shutdown pathclient cancel = stream drop
The pattern, not the product. Green = adjustable parts · vermilion = where a human decides.
Run Desk

Clean roomReconstruction of patterns I own. No Blackstraw code, data, prompts or clients.

Architecture: you, over SSE, to a supervisor that routes to 3 specialists. Specialists call scoped MCP tools. Reads go straight to the store; a propose-write stops at the approval gate, and only an approved proposal is written to the store. Record #42 is at rev 12.SSEpropose-writeapproved onlyreadyou · browsersupervisorroutes · holds no toolsA · readerreadB · analystreadC · editorread · proposeMCP server · scoped toolsevery call → audit row (digest)approval gate · younothing passes without youstore · record #42rev 12 · openlast writer: seed

Supervisor → 3 specialists → scoped tools. The supervisor and specialists are a team system I contributed to; the durable state with revisions, run cancellation and the MCP audit trail are my parts, checked by 28+ tests I wrote for authorization, stale writes and pagination.

1 · Give the supervisor a task

IDLEidle · give the supervisor a taskcheckpoint · step 0

2 · Approval gate

No write pending. Specialists hold no write scope: every propose-write stops here.

3 · Two ways to stop, one handler

stops the run, drops any proposal not yet written, keeps the checkpoint

4 · Another writer

No locks. A pending write carries the revision it read; the store writes only if that revision still holds, else the run re-reads and proposes again.

Run trace

  1. Nothing yet. Pick a task above.

5 · MCP audit log

Each call keeps tool, scope, latency and a SHA-256 of its arguments. Raw arguments are never stored.

No MCP calls yet. Every specialist tool call lands here.

6 · Check a digest

Start a run, then pick an audit row: you can edit the arguments it was made from and watch the digest stop matching.

idle · give the supervisor a task

Reconstruction: patterns from my production work, rebuilt from scratch with simulated timings
Show the numbers as a table
Run trace · idle · give the supervisor a task
StepActorToolScopeSim latency (ms)What happened
No steps yet.
MCP audit log (arguments are not stored)
#StepAgentToolScopeSim latency (ms)Args digest (SHA-256)
No MCP calls yet.
Record #42 (durable, survives runs)
RevisionStatusLast writerApproved writes committedCheckpoint (step)
12openseed00

Read the case study →

Sheet 05 / 08 · Graph Hacks 2026 · Oct 2026

Evidence

In short: Why-Graph: agents must prove why production broke as a graph path, then block the repeat.

Event
Graph Hacks 2026 (WeMakeDevs × FalkorDB), 15–18 Oct 2026
Team
Solaimuthu + Shoban Ravichandran
My part
Agents, MCP server, PR guard, UI, demo app, deploy
Checked by
167 passing tests
Result
judging pending
Sheet
05 / 08

AI coding agents now ship around the clock, so they also break production around the clock. And when an LLM explains an outage, you usually can't check its answer.

Why-Graph is what Shoban Ravichandran and I built for Graph Hacks 2026 in about thirty hours. Four agents investigate an incident on a shared FalkorDB graph, and every claim they make must cite a named graph query. Before a finding is recorded, the query is re-run; if the nodes and edges aren't really there, the claim is rejected. Confidence comes from the graph, never from the model.

Then the system learns. A confirmed incident becomes a lesson and a detector, and both are exposed to other coding agents as MCP tools and as a pull-request guard. The next agent that tries the same mistake on a different service gets stopped, with the reason and the evidence attached.

I built the agents, the MCP server, the PR guard and the UI; Shoban built the graph and memory layer. The incident below was staged on our astro-shop demo app, not a real company's production. Judging is still pending.

Solaimuthu's pencil drawing of a scorpion, construction lines visible.The same scorpion's silhouette redrawn as a pencil network graph: Why-Graph's seed services and their calls, with the raised tail as a four-step evidence path.
The bug becomes the graph: my scorpion sketch, and Why-Graph's real seed services laid out in its silhouette.
Try to break shipping
frontendcartcatalogcheckoutshippingpaymentp99 446 msquotetimeouts_ms.paymentconfig key on checkoutcommit b9f6186PR #1 · AI agenttimeout800 → 200 mstimeout600 ms
  • what happened
  • evidence, and what you can change
  • a verdict, always with a word

Act 1 of 5

An AI agent ships a timeout cut

In the live run, the team's AI coding agent wg-chaos opened PR #1 on the astro-shop demo app to “reduce checkout latency”. It cut checkout's timeout on payment from 800 ms to 200 ms. Payment's p99 at the time was 446 ms.

wg-reviewer asks Why-Graph: check_change_risk

No known risk

“No known risk: no lesson or detector matches this change.”

That answer was honest. The seeded history deliberately leaves this failure pattern out, so the graph had never seen it and no detector could fire. The PR merged.

Team build with Shoban Ravichandran at Graph Hacks 2026 (judging pending). Shoban built the seed graph, query catalog, lessons and detectors; Solai built the agents, PR Guard and MCP tools, the astro-shop demo app and its simulator.

Act 1 of 5. PR #1 cuts the checkout to payment timeout from 800 to 200 ms, below payment p99 446 ms. Guard: no known risk. Merged.

Simulation using recorded numbers from the Why-Graph repo (README, scenarios/act2.md, lessons.py)
Show the numbers as a table
Recorded numbers behind the Why-Graph guard simulation, and the current simulated verdict
WhatValueSource
checkout → payment timeout before PR #1800 mshttps://github.com/redwing-381/astro-shop/pull/1
checkout → payment timeout after PR #1200 mshttps://github.com/redwing-381/astro-shop/pull/1
payment p99 at PR #1446 msredwing-381/we-make-devs-hack · README
checkout error rate, staged INC-103162%redwing-381/we-make-devs-hack · README (staged on the astro-shop demo app)
frontend error rate, staged INC-103141%redwing-381/we-make-devs-hack · README
root_alert query11.0 msredwing-381/we-make-devs-hack · README (query catalog table)
blast_radius query4.3 msredwing-381/we-make-devs-hack · README (query catalog table)
suspect_changes query10.6 msredwing-381/we-make-devs-hack · README (query catalog table)
config_vs_reality query16.9 msredwing-381/we-make-devs-hack · README (query catalog)
L-1 factor1.5×redwing-381/we-make-devs-hack · whygraph/memory/lessons.py
D-1 backtest1 of 39redwing-381/we-make-devs-hack · README
shipping → quote timeout before600 msredwing-381/we-make-devs-hack · scenarios/act2.md
shipping → quote timeout the agent proposed100 msredwing-381/we-make-devs-hack · scenarios/act2.md
quote p99, seed baseline300 msredwing-381/we-make-devs-hack · seed/data/core.json (CALLS svc:shipping→svc:quote)
quote p99 at the recorded Act 2 run≈302 ms(derived)redwing-381/we-make-devs-hack · README (safe_min 453) + whygraph/memory/lessons.py (ceil 1.5×p99)
safe_min returned when Act 2 was blocked453 msredwing-381/we-make-devs-hack · README; scenarios/act2.md
Your shipping → quote timeout600 msthis simulation
safe_min computed here, ceil(1.5 × quote p99)453 msthis simulation
Guard verdictOK · within L-1this simulation · after L-1

Read the case study →

Sheet 06 / 08 · Hacker House Goa 2026 · Aug 2026

Budget

In short: RAG in Goa: retrieval inside a hard 200 ms deadline. We measured a reranker and cut it.

Event
Hacker House Goa 2026 task, Aug 2026
Team
Team McQuade: Solaimuthu, Merlyn Michelle, Shoban Ravichandran
My part
Retrieval rework, GPU index, voice output, latency + architecture docs, deploy
Index
215,608 chunks
Retrieval
P95 76.85 ms live
Sheet
06 / 08

RAG in Goa answers spoken questions in Hindi, Bengali, Tamil or Marathi. Speech becomes an English query, retrieval finds the passages, a model writes an answer that must cite them, and the answer is spoken back in your language.

The brief cared about latency, so retrieval runs against a hard 200 ms deadline. If time is running out, a stage is skipped and logged as a degradation instead of blowing the budget.

The decision I'm proudest of is one we didn't ship. A cross-encoder reranker improved the ranking, so we measured it across the benchmark set. It ate about ~160 ms of the 200 ms. We left it switched off. On this team, a refusal beats a fluent miss, and an honest number beats a nice one: the latency claim covers retrieval only, and the speech and model time are shown separately.

We were a team of three. I reworked the retrieval pipeline, built the GPU index and the Bulbul voice output, and wrote the latency and architecture docs. Merlyn built the API and the guardrails; Shoban built the voice UI and ran the benchmark.

Retrieval budget race

Retrieval against the deadline

Machine
Percentile
Reranker

Live · Fly.io Singapore, rerank off as shipped: embed → HNSW → BM25 → fusion → context expand, 400 queries after warmup.

  • embed (ONNX int8)33.49 ms
  • search: HNSW · BM25 · fusion · expand43.36 ms(total − embed, derived)
  • total76.85 ms= 38.4% of the budget

LATENCY.md prints search P95 as 55.41 ms. Each percentile is a different query, so stage percentiles don’t add up to the total; the bar draws total − embed, which is how benchmark.py defines search.

Outside the budget

One live request at P50, end to end. The 200 ms claim covers retrieval only; speech and the LLM are someone else’s servers, shown in grey.

speech-to-textvoice only · not timed
  • retrieval 47.5 ms live P50, inside the budget
  • LLM ~617 ms someone else's GPU, reported on its own
  • network hop and the rest ~436 ms (derived)
  • = live HTTP P50 ~1.1 s
  • speech-to-text: not timed in LATENCY.md; the benchmark sends typed text, so it is in neither the table nor the ~1.1 s.

Gates · illustrative trace

The path the pipeline code gives each question. Not a recorded run: outcomes are the ones the README states.

Sample questions (README)
Traps

Question · English · what type of mountain is Mount Fuji

  1. 1. Input gatepassnot unsafe, not an injection, readable
  2. 2. Translate to Englishnot neededalready English
  3. 3. Retrievalranruns under the deadline
  4. 4. Domain gatepassthe corpus has close passages
  5. 5. LLMranmay only speak from the retrieved passages
  6. 6. Output gatecheckschecks citations, then groundedness
  7. 7. Answerrancomes back in English
Passes both gates before the LLM; the answer comes back in English.

Deadline drill

Before each optional stage the retriever checks the clock and skips the stage, logging it, if too little time is left. Lower the deadline to see what a request with these P95 times (Live · Fly.io Singapore) would give up.

200 ms
  1. 1. HNSW ef_search · needs 140 ms leftruns

    Checked at 33.49 ms, before its cut-off at 60 ms.

  2. 2. BM25 leg · needs 140 ms leftcan’t tell

    Checked between 33.49 and 76.85 ms, cut-off 60 ms. LATENCY.md doesn’t split dense from BM25 time, so the check could land on either side.

  3. 3. Rerank · needs 120 ms leftnot loaded

    No reranker in the shipped config (RAGOA_RERANK=0), so there is no check.

Grey bar: when the check runs, as far as the recorded numbers pin it down. Dashed green: its cut-off (deadline − reserve). Vermilion: the deadline.

trace.degradations = [] · maybe "sparse_skipped"

ef_search 64 · finishes inside the deadline

team of 3 · Solai: retrieval-pipeline rework · Merlyn: API and guardrails · Shoban: voice UI and the benchmark

Live · Fly.io Singapore, P95: retrieval 76.85 ms of the 200 ms budget, 38.4%, under budget. Rerank off, as shipped.

Simulation using recorded numbers from LATENCY.md in the RAG in Goa repo
Show the numbers as a table
Retrieval latency, rerank off (ms, LATENCY.md)
MachinePctembedsearchtotaltotal − embed (derived)of 200 ms
Local · MacBook Air M2P502.214.997.335.123.7% · under
Local · MacBook Air M2P953.138.7111.148.015.6% · under
Local · MacBook Air M2P993.9410.9413.659.716.8% · under
Live · Fly.io SingaporeP5018.6125.2147.528.8923.8% · under
Live · Fly.io SingaporeP9533.4955.4176.8543.3638.4% · under
Live · Fly.io SingaporeP9941.2462.9688.847.5644.4% · under
Rerank on (earlier local run)
total P50165 ms
total P95197 ms
total P99not recorded
rerank alone~160 ms
live with reranknot recorded (live ran with rerank off)
A live request at P50: only retrieval is inside the budget
speech-to-textnot timed
retrieval47.5 ms
LLM~617 ms
network hop and the rest (derived)~436 ms
live HTTP P50~1.1 s
Deadline checks (ragoa/index/retriever.py) at 200 ms, current row
Checkneeds leftcut-offchecked atoutcome
HNSW ef_search140 ms60 ms33.49 msruns
BM25 leg140 ms60 ms33.49–76.85 mscan’t tell
Rerank120 ms80 msn/anot loaded

Reserves from ragoa/config.py: BM25 leg and ef 140 ms, rerank 120 ms.

Read the case study →

Sheet 07 / 08 · Everywhere · 2024–2026

Loose Pages

In short: Just start: a dated ledger of hackathon builds, and ProjectX, the one I built for myself.

Builds
25 in the ledger, newest first
Odoo 2026
Finalist: EcoSphere qualified the team; Tally Stack was built at the finals
Rule
No result recorded means no result claimed
Sheet
07 / 08

Most of what I've built started the same way: a weekend, a problem statement, and the decision to just start. Here is every one of them, dated, with who I built it with and what actually happened. Where there was no result, it says so.

One of them I built only for myself. I wanted to carry a keypad phone to the office without missing the messages that matter, but chat apps have no API for personal messages and the office machines blocked my email. So ProjectX reads the phone's notifications directly, decides what's urgent, and texts only that to the keypad phone.

ProjectX: three gates to a keypad phone
The three gates8 notifications in. Gate 1, VIP sender, caught 1. Gate 2, keywords, caught 2. 5 reached gate 3, the AI gate; with illustrative verdicts 1 would be forwarded. 4 stay on the smartphone.8 notifications in1VIP senderinstantsender contains a VIPSelected message stopped here1 to SMS2Keywordsinstanttext contains a keyword2 to SMS3AI gate~2 sGPT-4o-mini, not run here1 to SMS4 stay on the smartphone

Gates 1 and 2 are the original's rules, run here exactly. Gate 3 is not run in your browser: its time (~2 s) is self-reported and its verdicts here are illustrative presets.

Selected · Chat · Mom

“Are you eating properly? Call me tonight.”

  1. gate 1caught: the sender contains “mom”
  2. gate 2not reached
  3. gate 3not reached
Urgent · SMS sent
ProjectXSMS 1/4
CHAT: Mom - Are you eating properly? Call me tonight.
53/160 chars

the keypad phone in the pocket

Incoming · 8 notifications · pick one to trace it

“AI:” marks an illustrative verdict

Gate 1 · VIP senders

Caught when the sender contains one of these, ignoring case.

Gate 2 · Keywords

Caught when the text contains one, ignoring case. Like the original it is a substring match, not a whole word, so “important” also catches “unimportant”.

Send your own notification

Gates 1 and 2 route it instantly. If it reaches the AI gate, nothing is sent here.

Would stop at: type something to see the route.

CLI on PyPIpip install projectx-cli

17 property tests (Hypothesis)

Simulation using recorded numbers from the ProjectX repo (classifier and SMS formatter)
Show the numbers as a table
VIP senders: mom, dad. Keywords: urgent, asap, deadline, emergency, important. SMS limit 160 characters. Gate 3 verdicts are illustrative, not model output.
#AppSenderStopped atWhyVerdictSMS chars
1ChatMomgate 1, VIP senderVIP sender: momurgent, SMS53
2EmailDealsgate 3, AI gatePromotional emailnot urgent, stays (illustrative)none
3EmailIT Supportgate 2, KeywordsUrgent keyword: urgenturgent, SMS57
4Team chatStandup botgate 3, AI gateAutomated digest that can waitnot urgent, stays (illustrative)none
5ChatCousingate 3, AI gateFamily health news asking for a callurgent, SMS (illustrative)80
6CommunityMeetup groupgate 2, KeywordsContains urgent keyword: importanturgent, SMS83
7Team chatPriyagate 3, AI gateReview request with no deadlinenot urgent, stays (illustrative)none
8CommunityStudy groupgate 3, AI gateCasual group chatnot urgent, stays (illustrative)none

The ledger

  1. Oct 2026

    Why-GraphGraph Hacks 2026 (WeMakeDevs × FalkorDB)

    Agents that must prove why production broke as a graph path, then block the next agent from repeating it.

    team of 2 · with Shoban Ravichandran

    453 ms

    RepoDemo videoArchitecture walkthrough

    Result: judging pending

  2. Oct 2026

    Dark FactoryWeAreDevelopers × BAND hackathon

    He wrote the rules for the AI team, not the app, then debugged the team like a production system.

    solo

    120/120

    Repo

    Result: result not recorded

  3. Oct 2026

    Statistics Unit 13

    A single-file explorable that answers why the grouped mean is only an estimate.

    Result: teaching material

  4. Sep 2026

    Tally StackOdoo Hackathon 2026 finals (24 hours)

    An accounting product on Odoo 17 where every AI-proposed write waits for a human to confirm it.

    team of 3 · with Merlyn Michelle, Shoban Ravichandran

    24 hours

    Repo

    Result: built at the finals; result not recorded

  5. Aug 2026

    RAG in GoaHacker House Goa 2026 task

    Multilingual voice Q&A over 215K chunks with retrieval inside a hard 200 ms deadline.

    team of 3 · with Merlyn Michelle, Shoban Ravichandran

    76.85 ms

    RepoLive UI (cold-start)LATENCY.md

    Result: no result recorded

  6. Jul 2026

    EcoSphereOdoo Hackathon 2026, virtual preliminary round

    An ESG platform with a live 0–100 score and a What-If simulator. It took the team to the finals.

    team of 3 · with Shoban Ravichandran, Merlyn Michelle

    top 10%

    RepoLive demo

    Result: qualified for the finals (top 10% of 20,000+)

  7. Feb 2026

    AgentPayETHGlobal HackMoney 2026

    Six AI agents debate and pay each other per round over state channels: 20+ payments, 2 on-chain transactions.

    Result: no result recorded

  8. Jan 2026

    ProjectXDynamous × Kiro hackathon

    Texts only the truly urgent messages to a keypad phone, so the smartphone can stay in the bag.

    Result: no result recorded

  9. Dec 2025

    MomentAI Partner Catalyst 2025 (Confluent challenge)

    Streams employee actions through Kafka; rules decide the obvious cases and Gemini only the grey zone.

    solo

    0.3–0.8

    Repo

    Result: no result recorded

  10. Dec 2025

    Pick For MeYelp AI API hackathon

    An AI that commits to one choice instead of handing you fifty.

    solo

    Repo

    Result: no result recorded

  11. Dec 2025

    DeepFocus Relay

    The same email-triage problem built three ways: CrewAI, AutoGen and a hybrid.

    solo

    Repo

    Result: personal project

  12. Dec 2025

    Mirmer AIKiroween 2025

    Several LLMs answer, review each other anonymously, and a chairman model writes the final answer.

    Result: no result recorded

Every build, with my part in each →

Sheet 08 / 08 · Here · now

Your Turn

In short: Ask the sketchbook anything about my work, or leave a note.

Based in
Chennai, India
Sheet
08 / 08

That's the sketchbook so far. If you want to talk about agents, retrieval, or anything on these pages, write to me. I read everything.

Pencil drawing of a small bird landing on a branch with both wings fanned wide, construction grid faintly visible.

Back where it started. The original drawing, unmodified.

Ask the sketchbook

Answers come only from this site. Code checks every claim before it’s inked in vermilion; anything that fails is struck, with the reason.

Leave a note

Sent once, to me only. See privacy.