Skip to content
Sailesh Dahal

AI Integration Consultant

I'm Sailesh.I help teams put AIinto production.

Most AI projects die between the demo and the deploy. I work on the part after the demo: the retrieval that has to be right, the evals that gate the release, the guardrails and the audit trail, the cost per request.

  • Regulated compliance review
  • Customer operations on WhatsApp
  • Wearable-backed health assistant
Start a conversationSee what I've built

The offer

How I work with teams

Four shapes. Most engagements start with the first.

  1. 01AI opportunity auditTwo weeks inside your product and your team’s workflows. You get a ranked list of where AI actually pays, with the cost, the failure modes and the eval you would need for each, plus an honest list of where it does not.
  2. 02Build the first integrationOne workflow, taken all the way to production: retrieval, prompts, tool calls, an eval suite that gates deploys, cost and latency budgets, and a human-in-the-loop path for the cases the model should not decide alone.
  3. 03Agents and internal workflowsAgents that touch your real systems through typed tools and MCP servers, with permissions, audit trails and a kill switch. Built so the interesting part is your data, not the plumbing.
  4. 04Make your product legible to AIStructured data, markdown twins, llms.txt and in-page agent tools, so assistants describe your product correctly instead of guessing. This site is the reference implementation.

Want the concrete version? RAG, agents, evals, self-hosted inference. Example briefs, cheapest to deepest

Example implementations

Open any of these and check it. Nothing here is a screenshot of a prototype.

Products I've built

Shipped, live and open to inspection. Entries marked LaunchBox were built at the studio where I am lead engineer; the rest I built outside it.

What I build

The five hard parts

Where AI projects actually fail, and the work behind each one. Counts are merged pull requests.

Contract record since February 2021: 32 contracts at 100% job success, one of them running 4,494 of the 4,963 hours. Open the profile and check it ↗

  1. 01

    Evals, observability and guardrails

    The part that decides whether a model change ships or gets reverted. Most teams wire up an LLM and then have no way to tell whether last week's prompt edit made it worse.

    • Rayu: the eval printer and the memory-eval accounting, plus a live end-to-end probe that runs against the real model and database instead of a mock.
    • Thesis: releases gated on word error rate, a model registry carrying checkpoint provenance, and a reporting loop that closes the delta between two runs.
    • This site: a build gate that fails the deploy when a page loses its markdown twin or an image inside one stops resolving.

  2. 02

    Memory and personalisation

    Getting a model to remember the right thing and retrieve it for the right reason, which is harder than storing the conversation.

    • Rayu: routing a fact by its topic rather than its wording, so a memory written one way is still found when asked another way.
    • Grounded in live glucose, sleep, HRV and cycle data, all end-to-end encrypted, on a wearable SDK I integrated.

  3. 03

    Retrieval and document AI

    Answers that cite the passage they came from, so a reader can disagree with the machine instead of trusting it.

    • Solas: ingest the policy documents, extract enforceable rules, then review copy against them with every verdict linked back to its clause. 78 merged PRs on the frontend, 65 on the backend.
    • ReceiptSasa: receipt and invoice extraction, OCR into an OpenAI parsing step and out to QuickBooks. My PRs there are the OCR and the accounting export.
    • Thesis: retrieval over video transcripts, where the transcript itself is the unreliable part.

  4. 04

    Voice and multimodal

    Audio, video and documents through the same pipeline, with latency budgets that keep a conversation feeling like one.

    • Rayu: live transcripts during a call and the summary generated after it.
    • Code-switched Nepali and English speech recognition, LoRA fine-tuned with gradient checkpointing so the run fits on a single T4.
    • Solas: the same review pass over images, video and audio, not only text.

  5. 05

    Agents and tool use

    Agents that act on real systems through typed tools, with permissions, an audit trail and a kill switch.

    • Palete: the backend and the integration layer behind agents running customer operations, wired into WhatsApp, Shopify, HubSpot and Zendesk.
    • This site: read-only WebMCP tools registered inline, so an in-browser agent queries it through a typed contract instead of scraping the DOM.
    • 220 merged pull request reviews across one organisation, which is the part of technical direction that leaves a record.

Recent writing

All 11 posts

Series

Open source

Repositories I wrote.

Ordered by what I still touch, not by stars. The busiest one is missing: 586 commits this year went to the thesis project, code-switched Nepali and English speech recognition, still private until it is published.

Merged into other people's projects

I help teams put AI into production.

Start a conversationhi@saileshdahal.com.np