$ open ./projects → fin-research-agent/ · yourspeakreps/ · iamrishavraj1.com · résumé

Hi, I'm Rishav — a full-stack engineer who got tired of demos that die the moment real data shows up.

By day I build and run production services: Next.js + TypeScript frontends, Python backends. The rest of the time I'm retooling into AI engineering — properly, not tutorial-style: retrieval quality you can measure, agents that cite their sources, and evals that block a bad PR before it merges.

I like problems where the hard part is honesty. A 300-page SEC filing that almost answers your question is more dangerous than one that clearly doesn't — so my projects are built around knowing the difference.

If my work saved you an evening — a chai says it best. Pay what it's worth.

☕ Buy me a chai	🔗 Find me
scan with any UPI app · rishavraj2000@yescred

Ask "Which company grew revenue fastest after its 2023 acquisition?" → get a cited answer, or an honest "the filings don't say."

flowchart LR
    E["SEC EDGAR<br/>10-K · 10-Q · 8-K"] --> I["Ingestion<br/>parse → chunk → embed"]
    I --> DB[("Postgres<br/>+ pgvector")]
    DB --> H["Hybrid retrieval<br/>BM25 + vector → rerank"]
    DB --> Q["SQL tool<br/>structured financials"]
    H --> A["Agent"]
    Q --> A
    A --> O["Answer with citations<br/>or honest refusal"]
    V["Evals — golden set<br/>recall@k · MRR · CI-gated"] -.-> A

• Hybrid retrieval (BM25 + pgvector) — measured with recall@k and MRR, not vibes
• A raw, no-framework tool-using agent: document search, SQL over financials, calculator
• An eval suite that gates CI — the part most RAG repos skip
• Architecture, docs and eval design are up; code lands week by week

Speaks questions out loud (Ollama · qwen3:14b), listens to spoken answers (faster-whisper), and coaches: score /10, filler-word counter, STAR-mode feedback. ₹0 per session, works offline. The interesting bit is the design — the server runs the interview (question order, when to push), the LLM only converses. Deterministic control, natural feel.

• Measure, then ship. Every retrieval or prompt change runs the evals.
• No framework magic in the core loop. If I can't explain a line in an interview, it doesn't ship.
• Boring code, interesting products — in that order.

• Dharmveer Award — High Integrity & Ownership, for delivery quality and scalable architecture
• Pull Shark ×2 · YOLO — GitHub achievements for merged contributions
• 15,000+ reads across technical writing on dev.to, Hashnode and Medium

Frontend Next.js · React · TypeScript · Tailwind  |  Backend Python · FastAPI · Postgres + pgvector  |  Ops Docker · GitHub Actions  |  AI Ollama · OpenAI-compatible APIs · faster-whisper

Built by hand — no framework for the profile, Next.js for the site.