Bravim Purohit
M.S. Computer Science · Illinois Institute of Technology · May 2026
I build AI systems that hold up in the real world — ambient scribes, agentic pipelines, fine-tuned clinical LLMs, and the consensus, queueing, and serving infrastructure underneath them. Three years of shipping in healthcare, where bugs have real consequences and good evaluation is non-negotiable. Most recently the sole engineer on an enterprise healthcare deployment at Kinetic Systems in SF. Open to founding-engineer and ML roles at applied AI startups where the work still matters.
“The best way to predict the future is to invent it.”
— Alan Kay
Compact summary. Detailed engagement notes on request, in interviews.
Three things on my plate right now.
Sole engineer on an enterprise deployment in a regulated industry, from customer workflow through to production: agentic orchestration, retrieval and context design, document intake, and evaluation harnesses for systems that do not return the same answer twice.
Raft consensus in C++20, a three-broker task queue, a GraphRAG engine, a sandboxed code-repair agent, an inference gateway, a streaming orchestrator, and two forward-deployed systems. Each one has a spec, a benchmark, and an honest-claims register.
Can frontier AI agents actually handle the administrative work that burns out clinicians? HealthAdminBench (ICLR 2026, Stanford collab) is our attempt to measure that rigorously.
Eight of these landed in one month, across four disciplines — consensus and storage, agent infrastructure, inference serving, and forward-deployed data systems. Each carries a spec, a benchmark, and a register of what it does not prove. Company work is reserved for interviews, kept off the public site by design.
Real-time ambient scribe for clinical conversations. WebSocket speech-to-text for live transcription, incremental SOAP note generation, and a Risk Sentinel module that surfaces clinical risks with transcript-backed evidence. The LLM layer is deliberately non-blocking: if generation fails, transcription stays alive, because a crash must never interrupt a doctor mid-conversation.
Raft consensus implemented from scratch in C++20 — no etcd/raft, no NuRaft. Leader election, log replication, and crash recovery across a multi-node cluster, with linearizable reads via ReadIndex and leader-lease paths rather than routing every read through the log. The consensus core is deliberately pure: no sockets, no threads, no clock reads, so every scenario replays from a seed.
Background-job engine on a pluggable broker interface with three implementations — Redis Streams, Kafka, RabbitMQ — so the same workload is benchmarked across brokers rather than assumed to behave alike. Exactly-once effect — not exactly-once delivery — via at-least-once delivery plus idempotent commit plus fencing tokens that stop a partitioned worker from writing after its lease is reassigned.
Retrieval system that parses documents into a semantic knowledge graph instead of treating retrieval as top-k similarity. Every node and edge carries provenance back to its source span, so an answer traces to the exact text it came from. A query router picks graph traversal, vector search, or both, then fuses with Reciprocal Rank Fusion; a critique step can reject an answer and force re-retrieval, and abstention is a permitted outcome rather than a forced guess.
Agent that closes the loop on its own failures: writes code, runs it in a per-run Docker sandbox, parses the traceback back into structured context, and iterates until tests pass. Sandboxing is the load-bearing part — no network, read-only root, dropped capabilities, and CPU/memory/process/wall-clock limits, because an agent running code it just wrote needs a containment boundary rather than a promise of good behaviour.
Gateway between client apps and self-hosted models: semantic caching, backend routing, rate limiting, and per-tenant accounting behind one OpenAI-compatible surface. The committed threshold sweep is a negative result, and it is the most useful thing in the repo — against adversarial near-misses (Australia vs. Austria, enable vs. disable), no single similarity threshold makes the cache both useful and safe.
Orchestration backend where a run outlives any single client connection and a reconnecting client recovers exactly what it missed. Every run appends to an immutable events table under a gap-free per-run sequence number allocated inside the run transaction — a hole in the sequence is indistinguishable from lost reasoning. Cross-instance fan-out via PostgreSQL LISTEN/NOTIFY over a transactional outbox, so no sticky sessions.
Action-taking helpdesk agent that drafts operational changes — ticket updates, status transitions, refunds — for a human to confirm before anything executes. The design commitment is a hard read/write split: the agent’s own tool surface is read-only over MCP, and every write goes through a separate executor that only acts on an HMAC-signed, single-use, expiring approval token. An agent with write credentials and a review step it can bypass is not human-in-the-loop.
Two tiers for turning messy client documents into strictly conformed data: an interactive tier for onboarding one client’s files, where an operator reviews every extracted field against its source span, and a batch Medallion lakehouse — Bronze, Silver, Gold on Delta Lake — for processing a corpus at volume. Per-field confidence is surfaced so review effort goes where the extractor is unsure.
Clinical retrieval-augmented generation pipeline combining text and imaging embeddings for case-aware diagnostic insights. AstraDB for vector retrieval, MedLLaMA for grounded report generation.
Pulled live from the GitHub API at build time. Last 52 weeks of activity.
A supporting credential for the applied work above. Full author lists and citations on arXiv and in the journal.
Co-author. Stanford University collaboration.
A benchmark evaluating frontier Computer-Use Agents (Claude, GPT) on real healthcare administration workflows: prior authorization, appeals, and DME order processing. The paper shows where current CUAs break on production healthcare admin and what closing the gap requires.
Peer-reviewed journal article.
An AI system for automated land use monitoring, detecting and classifying changes such as deforestation, urban sprawl and agricultural shift from satellite imagery. Awarded Best Engineering Innovation for the month of publication.