Start a conversation
Tell me what you are trying to do. If it is a bad idea I will say so — and if there is a cheaper way to get the same result without a model in the loop, I will say that too. No pitch deck, no discovery call before there is anything to discover.
- What you are trying to do, in a sentence or two.
- What exists today — a product, a pile of documents, a workflow, or nothing yet.
- What "working" would look like, and roughly when you need it.
Rather not use a form? Email hi@saileshdahal.com.np or find me on LinkedIn and X.
What I can help you build
Example briefs, ordered from the cheapest useful thing to the deep end — not a list of past projects. The work I have actually shipped is on the home page, with links you can open.
Start here
- AI where it already fitsSummarise, classify, extract, draft. Wire a model into the workflow your team already runs — a support macro, a triage queue, a report that writes its own first draft — with a prompt you can read and a test that tells you when it drifts.
- Ask your documentsRetrieval over your own corpus: contracts, policies, tickets, wiki, PDFs. Chunking and embeddings tuned to your documents, hybrid search with a reranker, and every answer citing the passage it came from so a reader can disagree with it.
Production
- RAG that survives contact with real dataThe version that holds up past the demo: incremental ingestion, freshness and permissions carried through the index, query rewriting, retrieval evals with real recall numbers, and a fallback that says "I do not know" instead of inventing a clause.
- Agents with handsAgents that act on your systems through typed tools and MCP servers — book, refund, update, escalate — with scoped permissions, idempotent actions, a full audit trail, and a kill switch. Human approval on anything expensive or irreversible.
- Evals, guardrails and the release gateA graded dataset built from your real traffic, an eval suite in CI that blocks a prompt or model change from shipping when quality drops, plus injection and PII checks, refusal behaviour, and per-request cost and latency budgets you can watch.
Deep end
- Your own inference stackWhen the API bill or the data boundary stops making sense: open-weight models served on your own GPUs with vLLM, quantisation and batching sized to your traffic, LoRA fine-tunes on your data, and a routing layer that sends the easy 80% to the small model.
- Multimodal and voiceVision, audio and documents in the same pipeline — read the invoice, watch the video, hear the call. Streaming speech in and out for a phone or in-app agent, with latency budgets that keep a conversation feeling like a conversation.
- AI inside a mobile appThe part most teams underestimate: streaming into a React Native or Flutter UI, on-device and hybrid inference, offline behaviour, token cost per session, and a review-safe story for what the model can and cannot do on a user’s phone.
None of these quite it? Describe it in your own words — the form takes free text for a reason.