SahilKhan

An AI review platform for GitHub and GitLab. Every pull request is cloned into a sealed sandbox, mapped with a code graph, scanned by static analyzers and investigated by tool-using LLM agents — then a judge model throws out anything it can't back with evidence. What survives lands on the PR as a walkthrough, inline comments with one-click fixes and a merge check. Ask it, and it writes the fix, the tests or the CI repair too.

• Agentic review pipeline

  Triage → plan → investigator agents that read files, walk callers and callees, and run shell commands inside the sandbox — each capped at 8 tool calls and 60k tokens.

• Built-in hallucination control

  A finding must sit inside the diff, apply cleanly as a patch and pass an LLM judge's confidence bar. Linter output is evidence for the agents, never posted raw.

• Sealed, disposable sandbox

  Read-only, non-root, no capabilities, 768 MB; the network is cut after the clone and the container is destroyed after. Repo text is fenced off as untrusted input against prompt injection.

• It writes the fix, too

  @hootpr autofix, generate unit tests (run and iterated until they pass), fix CI from the failed job logs, resolve merge conflicts — pushed as a commit or a stacked PR.

• Security suite

  Blast radius traces changed code back to the HTTP routes it can reach; a repo-wide attack-surface map and an on-demand security architecture review.

• Learns each team

  Preferences taught in the PR thread are embedded in pgvector and recalled on the next review. Linked repos, remote MCP servers and web search add context.

• Dashboard with a full trace

  Every review's stages, agent steps, LLM calls, tool runs and judge verdicts; a Monaco diff workspace with findings inline; analytics, reports, audit log and a public REST API.

• Measured, not guessed

  An eval suite of seeded bugs, reversed CVE fixes and clean PRs scores precision, recall and cost per PR, with ablations that switch stages off.

Type a US ticker and Assay reads the company's financials and SEC filings, computes every ratio in Python, benchmarks them against peers, and grades the business on profitability, financial health, growth, valuation and sentiment — then writes a cited investment memo with a BUY / HOLD / SELL call. The LLM writes the argument; it never produces a number and never picks the rating.

• Every number is verified

  Ratios are computed in pure Python. After the memo is written, every %, multiple and dollar figure is extracted and matched against those values — a transposed digit sends it back for a rewrite.

• A rubric makes the call

  27 metrics are scored 0–10 against named anchors and folded into five weighted dimensions. The model is told the rating and argues for it; if it disagrees, that goes in the caveats.

• Every claim is cited

  Qualitative sentences must cite a retrieved passage. Uncited claims, or citations to sources that don't exist, are removed before the memo is saved.

• Retrieval led by the numbers

  10-Ks streamed from SEC EDGAR plus news, chunked into pgvector. When revenue fell or leverage is high, it searches specifically for demand weakness or refinancing risk.

• Peer benchmarking

  Each metric gets a percentile against comparable companies — “operating margin 31%, better than 88% of peers” — with meaningless multiples filtered out.

• “What would make it a BUY?”

  The rubric runs backwards: a Scenario tab solves the smallest move in each metric that flips the rating, with draggable levers — no data fetch and no model call.

• Live pipeline, full trace

  Server-sent events stream each node as it finishes; the trace tab shows tokens, cost and citation rate for every run.

• A complete product

  Accounts with Google sign-in, compare, watchlist, PDF export, public share links, credits via Razorpay (test mode), quotas and a daily LLM budget guard.