# Solaimuthu A (Solai) > Software Development Engineer II at Blackstraw.ai, Chennai, India. I build multi-agent systems that show their work. Every number on the site comes from a sourced facts ledger. Team projects name the team and my part. ## Case studies - [Why-Graph](https://solaimuthu.vercel.app/work/why-graph): Agents that must prove why production broke as a graph path, then block the next agent from repeating it. (15–18 Oct 2026; my part: Agent runtime and the four agents, ingest, MCP server, PR Guard, UI, demo app, live deploy, demo video; result: judging pending). Verified: 453 ms (safe minimum the guard enforced); 167 (passing tests); 5 (MCP tools for other agents); ~950 (nodes in the demo graph). - [RAG in Goa](https://solaimuthu.vercel.app/work/rag-in-goa): Multilingual voice Q&A with hybrid retrieval inside a hard 200 ms deadline, and a reranker we measured and cut. (Aug 2026; my part: Retrieval pipeline rework, Sarvam Bulbul TTS, Colab GPU index, latency and architecture docs, deploy, mobile UI; result: no result recorded). Verified: 76.85 ms (retrieval P95, live); 215,608 (chunks indexed); 200 ms (hard retrieval deadline); ~160 ms (reranker cost, so it shipped off). - [Production agents at Blackstraw.ai](https://solaimuthu.vercel.app/work/production-agents): A supervisor, specialist agents, scoped tools, durable state, and a human before every write. (Aug 2025–now; my part: Persistence layer, AG-UI adapter, run cancellation, MCP audit trail, tests (contributor to the orchestration system); result: in production). Verified: 3 (specialist agents behind the supervisor); 28+ (tests I wrote). - [ProjectX](https://solaimuthu.vercel.app/work/projectx): Texts only the truly urgent messages to a keypad phone, so the smartphone can stay in the bag. (Jan 2026; my part: Solo: FastAPI backend, Kotlin Android listener, 3-stage classifier, PyPI CLI; result: no result recorded). Verified: 17 (property-based tests); 18 days (solo build); 160 characters (SMS limit). - [Hybrid ATR 72-600](https://solaimuthu.vercel.app/work/hybrid-atr): A student-team design study re-engineering a regional airliner around hydrogen fuel cells and batteries. (Jun–Jul 2025; my part: Hydrogen technology: fuel-cell design, hydrogen storage, safety systems; result: student design study). Verified: 3,258.1 kWh (mission energy); 22,500 kg (team's takeoff-weight target); 3 h 05 min (mission flight time). - [Tally Stack](https://solaimuthu.vercel.app/work/tally-stack): An accounting product built in 24 hours at the Odoo finals; the AI proposes, a human confirms. (5–6 Sep 2026; my part: The whole Next.js frontend, the pricing-anomaly agent, PDF reports, Odoo scaffolding (base assistant: Merlyn); result: built at the finals; result not recorded). Verified: 24 hours (build window at the finals); top 10% (how the team qualified). ## Pages - [Fair copy](https://solaimuthu.vercel.app/fair): the two-minute recruiter version - [Log](https://solaimuthu.vercel.app/log): every build, dated, with an honest result column - [Notes](https://solaimuthu.vercel.app/notes): short notes on decisions - [For agents](https://solaimuthu.vercel.app/agents): read-only MCP server at https://solaimuthu.vercel.app/mcp (claude mcp add --transport http solai https://solaimuthu.vercel.app/mcp) - Markdown twins: append .md to any page, e.g. https://solaimuthu.vercel.app/work/why-graph.md ## Builds - 2026-10 · Why-Graph (Graph Hacks 2026 (WeMakeDevs × FalkorDB)): Agents that must prove why production broke as a graph path, then block the next agent from repeating it. Result: judging pending. - 2026-10 · Dark Factory (WeAreDevelopers × BAND hackathon): He wrote the rules for the AI team, not the app, then debugged the team like a production system. Result: result not recorded. - 2026-10 · Statistics Unit 13: A single-file explorable that answers why the grouped mean is only an estimate. Result: teaching material. - 2026-09 · Tally Stack (Odoo Hackathon 2026 finals (24 hours)): An accounting product on Odoo 17 where every AI-proposed write waits for a human to confirm it. Result: built at the finals; result not recorded. - 2026-08 · RAG in Goa (Hacker House Goa 2026 task): Multilingual voice Q&A over 215K chunks with retrieval inside a hard 200 ms deadline. Result: no result recorded. - 2026-07 · EcoSphere (Odoo 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. Result: qualified for the finals (top 10% of 20,000+). - 2026-02 · AgentPay (ETHGlobal 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. - 2026-01 · ProjectX (Dynamous × Kiro hackathon): Texts only the truly urgent messages to a keypad phone, so the smartphone can stay in the bag. Result: no result recorded. - 2025-12 · Moment (AI Partner Catalyst 2025 (Confluent challenge)): Streams employee actions through Kafka; rules decide the obvious cases and Gemini only the grey zone. Result: no result recorded. - 2025-12 · Pick For Me (Yelp AI API hackathon): An AI that commits to one choice instead of handing you fifty. Result: no result recorded. - 2025-12 · DeepFocus Relay: The same email-triage problem built three ways: CrewAI, AutoGen and a hybrid. Result: personal project. - 2025-12 · Mirmer AI (Kiroween 2025): Several LLMs answer, review each other anonymously, and a chairman model writes the final answer. Result: no result recorded. - 2025-07 · MurfKiddo (Murf Coding Challenge 3): A voice-first learning companion for children. Result: no result recorded. - 2025-07 · LubTub AI: An AI emergency assistant. Result: no result recorded. - 2025-07 · Hybrid ATR 72-600 (ICAM summer programme, Toulouse): A student-team study re-engineering a regional airliner around hydrogen fuel cells and batteries. Result: student design study. - 2025-05 · STEM Explorer: A Three.js maths game for students. Result: no result recorded. - 2025-05 · CivicFix (AlgoArena): Citizens report, contractors bid, and the reporter rates whether it was actually fixed. Result: no result recorded. - 2025-04 · Zhagaram: A small web home for Tamil culture: food, art, a quiz and a forum. Result: personal project. - 2025-03 · LifeChain (HackVerse 2025): Encrypted family capsules that open only when enough trustees approve on-chain. Result: no result recorded. - 2025-03 · RESQ: One live channel for citizens' SOS calls, rescue teams and government alerts. Result: no result recorded. - 2025-03 · Urbexium (ETH San Francisco · Coinbase Developer Platform track): A public grievance platform. Result: First Place. - 2025-02 · Conmigos: A browser extension that saves and restores context across AI chat tools. Result: no result recorded. - 2025-02 · EventLaunch: The audience's phones launched the event: enough people joined and the projector flipped. Result: used at a live event. - 2024-04 · See. Sense. Stop. (LICET Pattarai): A camera sees a person enter a danger zone and a microcontroller sounds the buzzer. Result: no result recorded. - 2024-03 · Smart Transportation (TechnoTeen LICET Hackathon): RFID-based bus ticketing. Result: Runner-up. ## Notes - [Graph-only mode: proving the graph decides](https://solaimuthu.vercel.app/notes/graph-only-mode): If you can switch the LLM off and every verdict stays the same, the graph isn't decoration. - [Agents propose, people dispose](https://solaimuthu.vercel.app/notes/agents-propose-people-dispose): Why every write in my agent systems is a proposal that waits for a human. - [We measured the reranker and cut it](https://solaimuthu.vercel.app/notes/we-cut-the-reranker): A better ranking that ate most of a 200 ms budget. The number made the decision. ## Contact - Email: solaimuthu006@gmail.com - GitHub: https://github.com/redwing-381 - LinkedIn: https://www.linkedin.com/in/solaimuthu-a-87191225a/