I build autonomous AI systems end-to-end. Right now all-in on Propab.

An autonomous scientific research system. The goal is to compress humanity's discovery timeline: research that runs continuously, at machine scale, instead of in human-paced cycles.

I've been building this for a year now, including 4-6x full rebuild after internal audit showed me the architecture was wrong. under the hood:)

• Campaign orchestration. A persistent loop engine. Orchestrator and subagents are the same agent class at different hierarchy levels, native tool calls, no hardcoded intent parsing.
• Hypothesis trees, not flat search. Every synthesis output anchors to a completed parent node, so the search space converges instead of sprawling.
• Evidence binding. Every claim traces to a content-addressed artifact. Fabricated citations get caught at write time, not review time.
• Literature retrieval at scale. Hybrid BM25 + dense retrieval over open scholarly corpora (OpenAlex, Semantic Scholar), running on self-hosted index infrastructure because managed vector DBs don't survive contact with this scale.
• Sandboxed execution. Hypotheses get verified by actually running code, not by asking an LLM if they sound right.

Open-sourcing components as they stabilize.

• Languages: Python, TypeScript
• AI systems: multi-agent orchestration, RAG / hybrid retrieval, embeddings, sandboxed tool execution
• Data: PostgreSQL, pgvector, Redis
• Backend & infra: FastAPI, Node.js, Docker, Kubernetes, GCP, self-hosted dedicated servers
• Frontend: Next.js, React

Co-founded Accenox, a B2B software company for global clients. Team of 8, 30+ production systems shipped. Paused it to go full-time on Propab.

ps: I also contribute to open source repos (if you're maintaining one, ping me!)

X: @iykshani · LinkedIn: shani-singh1 · propabai.com