# Sailesh Dahal > AI integration consultant. I take LLMs, agents and retrieval from demo to production inside real products and real workflows, with evals, guardrails and traceability, not vibes. - Canonical: https://saileshdahal.com.np - Role: AI Integration Consultant - Contact: hi@saileshdahal.com.np - Posts: 12 - GitHub: https://github.com/saileshbro - X: https://twitter.com/sail_sail30 - LinkedIn: https://linkedin.com/in/saileshbro - Upwork: https://www.upwork.com/freelancers/~013e4f55757178f76f --- ## How I work with teams - **AI opportunity audit**: Two 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. - **Build the first integration**: One 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. - **Agents and internal workflows**: Agents 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. - **Make your product legible to AI**: Structured 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. ## Example implementations - [In-browser agent tools on this site](https://saileshdahal.com.np/llms.txt) (Agent interface): Read-only WebMCP tools registered inline, so an in-browser AI agent can query the site through a typed contract instead of scraping the DOM. Feature-detected, 1.1 KB gzipped, fetches nothing until a tool is actually called. - [Machine-readable content pipeline](https://saileshdahal.com.np/llms-full.txt) (Retrieval): Every page on the site serves a markdown twin at the same URL plus “.md”, indexed by llms.txt and llms-full.txt, with a build gate that fails the deploy if a twin goes missing or an image inside one stops resolving. - [Agent skills library](https://github.com/saileshbro/skills) (Agent tooling): A curated, versioned set of coding-agent skills following the skills.sh convention, the same packaging approach used to give a team’s agents a shared, reviewable playbook instead of ad-hoc prompts. ## Products I have built > Some were built at [LaunchBox](https://launchbox.tech), the studio where I am lead engineer; the rest outside it. Attribution is noted per product. - [Solas Compliance](https://solascompliance.com) (Policy → rules → review): Marketing compliance for finance and iGaming. Ingest the policy documents, extract enforceable rules from them, then review copy, images, video and audio against those rules, with every verdict linked back to the clause it came from, with a full audit trail. Built at LaunchBox. - [Rayu](https://rayu.ai) (Multimodal · memory · sensors): An agentic AI nutritionist paired with a 24/7 wearable, and the company I co-founded. Log a meal by photo or voice, and get advice grounded in live biometrics plus everything the model remembers about you, all end-to-end encrypted. Alpha since August 2026. - [Palete](https://palete.ai) (Agents · integrations): AI agents running customer operations, deployed in six African markets. My share is the backend and the integration layer they act through: WhatsApp, Shopify, HubSpot, Zendesk, and the payment rails those markets actually run on. Built at LaunchBox. - [Sambad](https://sambad.io) (Omnichannel inbox · iOS): One inbox for WhatsApp, Instagram, Facebook, email and website chat, with AI that replies, qualifies leads and routes conversations, wired into order systems, stock and CRM so it can act, not just answer. Apps: [iOS](https://apps.apple.com/np/app/sambad/id6758042230) - [Sajilo Sewa](https://www.sajilosewa.com) (Marketplace · iOS · Android): Nepal’s home-services marketplace with 50+ services across Kathmandu, Lalitpur, Bhaktapur and Pokhara. Customer app, provider app and the booking platform behind them. Apps: [iOS](https://apps.apple.com/np/app/sajilo-sewa/id1641579696) · [Android](https://play.google.com/store/apps/details?id=app.sms.smsuser) - [DayOf](https://dayof.net) (Scheduling · Flutter + Django): Wedding-day scheduling for the people running the day: build the running order, lock what cannot move, then share it as a link, text, image or PDF. I built the scheduling and sharing side over nearly two years, 40 merged pull requests and 124 commits. - [Flexera](https://joinflexera.com) (Hardware + app): Connected fitness: a Pilates reformer paired with a mobile app for instructor-led classes, with offline playback for a studio with bad wifi. Hardware, firmware boundary and app. Built at LaunchBox. ## What I build > On Upwork since February 2021: 32 contracts, 4,963 hours, $110K+ earned, 100% job success. One contract is 4,494 of those hours. Verifiable at https://www.upwork.com/freelancers/~013e4f55757178f76f ### 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. ### 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. ### 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. ### 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. ### 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 - [AI Week: the sandbox was never a sandbox](https://saileshdahal.com.np/ai-week-2026-08-09.md): Issue 2. Anthropic reviewed 141,006 evaluation runs and found three where Claude reached the real internet. One published malware to PyPI that ran on 15 real machines. OpenAI found two more with UK AISI and Irregular. The pattern is not rogue models, it is eval harnesses nobody threat-modelled. - [AI Week: open weights stopped being ideological](https://saileshdahal.com.np/ai-week-2026-07-30.md): Issue 1. Hugging Face's incident responders could not use commercial models to investigate their own breach, so they ran an open-weight model on their own hardware. Four days later, the strongest open-weight model ever released shipped its weights. Plus the alliance, the letter, and the $250 billion. - [Few takeaways on getting amazed with AI](https://saileshdahal.com.np/few-takeaways-on-getting-amazed-with-ai.md): On 24th Jan 2025, I attended an insightful session organized by Department of Artificial Intelligence at Kathmandu University (KU) . This session was by Dr. Saumendra Mohanty and was really insightful providing a different perspective on AI and it’s… - [Flutter FAQ: codegen and bundle size](https://saileshdahal.com.np/flutter-faq-codegen-and-bundle-size.md): After 2 years of not writing and ditching my readers, I am trying to get back to writing again. The past couple of years have been really busy with freelancing on Upwork, where I delivered multiple apps for clients that are generating steady MRR. I’m… - [🤳 Effortless Sharing: From external apps to your Flutter app in no time](https://saileshdahal.com.np/sharing-media-from-external-to-flutter-app.md): This comprehensive tutorial will utilize the share_handler package to quickly and easily share media from external applications to your Flutter app. ## Series - [AI Week](https://saileshdahal.com.np/series/ai-week.md): 2 parts - [🚀 Building a Fullstack App with dart_frog and Flutter in a Monorepo](https://saileshdahal.com.np/series/fullstack-dart.md): 6 parts ## More - [Contact](https://saileshdahal.com.np/contact.md): what I can build, and how to start - [All posts](https://saileshdahal.com.np/archive.md) - [Agent index](https://saileshdahal.com.np/llms.txt) - [Full corpus](https://saileshdahal.com.np/llms-full.txt)