AI Lab

Why I built it

Why do enterprise assistants answer confidently and wrong? Often because two sources quietly disagree, and retrieval grounds in whichever one it fetched.

What I learned

• Audit RAG by inverting it: extract atomic claims across sources, cluster them, and flag disagreement instead of answering.
• A grounding guardrail beats a clever prompt: no contradiction is reported without citations from two different documents.
• Suggested resolutions are labelled as suggestions, with the heuristic shown, never asserted as truth.

Why I built it

Climate tools show where it is hot or flood-prone. Could one prove the consequence of a fix before money is spent?

What I learned

• Prescriptive beats descriptive: the product is the simulation and the budget optimizer, not the map.
• Label modeled numbers honestly: every figure carries an uncertainty band, cited coefficients and its provenance.
• Heavy pipelines run at build time (satellite exports, model training), which keeps the runtime cost close to zero.

Why I built it

Can an agent swarm surface conflicting data for a person to reconcile, instead of hiding it behind a confidence score?

What I learned

• Replace opaque confidence scores with a plain Verified / Needs review badge and the number of sources.
• Make disagreement a first-class object: a conflict shows both sides, and one click records the canonical value.
• Meter every model run by actual token consumption, so cost is visible product state.

Why I built it

Most captioners fail by hallucinating or by crashing the grading harness. What does a captioner built for reliability look like?

What I learned

• Ground first: a facts-only scene report is the contract. If it is not visible, it does not exist.
• Judge with a different model family to avoid self-bias, and refine only when a caption misses the quality bar.
• Reliability is a feature: atomic writes, per-clip timeouts and deterministic fallbacks keep the output valid.

Why I built it

Where is a model genuinely necessary? The build answers task by task, and the interface badges which engine answered.

What I learned

• Use the model only where language needs reasoning (intent, register, translation); thresholds and routing stay deterministic.
• Every AI output carries a short, human-readable "why".
• Test the deployed app: Arabic and Hindi tokenize expensively, and a low token limit was silently truncating answers.

Why I built it

Can an agent do a recruiter's first screen honestly: cite proof for every match and show the gaps instead of hiding them?

What I learned

• The model perceives, code decides: the LLM reads the job description into a fixed skills vocabulary and writes the brief, while matching and scoring stay deterministic and explainable.
• Conservative ratings make the score credible. Every skill states its basis, and a role outside my record scores low.
• Every action is a draft the visitor approves: the email opens in their own mail app and the invite downloads as a file.

Why I built it

Can a portfolio answer a recruiter's specific question faster than a PDF, and do the next step for them, without inventing anything?

What I learned

• Constrain the model to a verified fact sheet and have it say "I don't know" outside it; let it propose actions only from an allowlist.
• Commands take a deterministic fast path; only open questions reach the model, which keeps answers fast and cheap.
• Hedge model calls across two providers: the first answer wins, and curated answers cover a full outage.