Two projects, built end to end
Built by Biresh during an internship at PGAGI Consultancy (Lead AI/ML Intern on SentLogic, AI/ML Intern on the eBay AI Arbitrage System), not as Biru Labs contracts.
One project for a client in the USA and one for a client in Estonia, both delivered by our founder through PGAGI Consultancy: AI, backend, integrations and billing. The source code is private, and we're happy to walk through the architecture on a call.
Chapter 01 · Instagram B2B AI DM Automation SaaS
Built by Biresh as a Lead AI/ML Intern at PGAGI Consultancy, not as a Biru Labs contract.
A B2B SaaS that automates Instagram DMs for merchants and creators. A merchant connects Instagram, and optionally Shopify, and AI handles customer conversations in the DMs. A separate AI Business Advisor helps merchants understand and run their own account.
Built by Biresh and the team during our internship at PGAGI Consultancy, with ownership across AI, backend systems, integrations, billing, knowledge management, and production infrastructure.
Instagram OAuth with encrypted tokens, account switching and a posts feed. Meta webhooks with signature checks handle comments and DMs: keyword triggers with exclusions, cooldowns and duplicate protection, then an automatic first DM.
We researched Instagram's messaging-window rules and adapted the flows around them, and passed Meta's app review after fixing the issues from the first submission.
Shopify OAuth, bulk catalogue sync, and webhooks for products, inventory and orders. Product edits re-embed the knowledge base automatically, and a nightly reconciliation job catches anything a webhook missed.
Orders are traced back to the Instagram conversation that started them, and a dashboard handles Shopify customer data requests (GDPR).
Claude Sonnet and Haiku on AWS Bedrock, Titan embeddings stored in pgvector, and query expansion for better retrieval. Simple questions go to the smaller model and knowledge-heavy ones to the larger. Bedrock Guardrails keep customer-facing and merchant-facing replies safe.
Each reply gets a confidence score that weights the evidence found (60%) above the model's own opinion of itself (40%). Below 0.7 the AI pauses and the merchant takes over. Complaints and business enquiries escalate automatically, with an email and a push notification.
It's a practical rule of thumb, not a statistical guarantee.
Merchants add URLs, PDFs, Word files, text, FAQs and product catalogues. Up to five versions, one live at a time, with rollback. The AI only reads the live version, and live documents can't be deleted by accident. A report suggests what's missing for the merchant's kind of business.
A floating assistant that answers "how are my automations doing?" from real database queries instead of guesses. Merchants can create, pause, edit or delete automations in plain English: similar names trigger a clarifying question, and destructive actions ask for confirmation.
It knows which page you're on, remembers past chats, searches the web for competitor questions and has its own daily AI budget.
Comment, story and DM-responder automations with follow-gates, email capture, images, buttons and tracked links. Unfinished automations are saved as drafts, and stopped ones go to an archive instead of being deleted.
A dashboard with pause and resume, human replies, and automatic pause with a grace period. Haiku classifies every message into seven types and tracks sentiment, and a nudge system follows up with people who go quiet mid-conversation.
Stripe checkout, portal and cancellation with idempotent webhooks, five tiers, and monthly conversation credits. Customer replies and the Advisor have separate daily AI budgets, so heavy customer traffic can't starve the merchant's own assistant.
A global kill switch puts the whole account in manual mode. Web Push works on Android and on iPhone home-screen web apps, with per-type preferences so merchants only get what they want.
Funnels, click-through, per-automation and per-product stats, peak hours, AI resolution rate, and link-click tracking with revenue attribution.
Role-based access, an append-only audit log, user and billing tools, a support inbox, a content CMS, broadcasts, platform analytics, two-factor login and time-limited read-only impersonation.
Confidence 0.72: the AI replies.
An illustration of the rule described above: evidence counts for 60% and the model's opinion of itself for 40%. Below 0.7 the AI pauses and the merchant takes over.
We chose cart links over UTM parameters and cookies, which can be stripped or blocked during checkout. Putting the attribution in the Shopify cart gave the backend a source of truth that survives.
A customer comments a keyword on a post.
The AI answers in the DM, using the merchant's knowledge base.
The reply carries a cart link with the conversation, workspace and channel.
The customer pays, and the attributes travel with the order.
The order webhook is verified and de-duplicated, then the sale is credited to the conversation.
Subscription status was cached in a login token that lasted seven to eight days, so a cancelled user kept paid access until they signed in again. We moved the check to the database.
We also added an automatic downgrade the moment a payment dispute opens.
In testing, the AI apologised to a customer who claimed to have found a cockroach in a sandwich the merchant doesn't even sell. Claims about products and complaints are now checked against the knowledge base before the AI admits or apologises for anything.
While a document was being re-chunked and re-embedded, the playground could answer from stale data, and the Advisor was reading drafts. We locked input during processing and limited retrieval to the live version.
Webhooks get retried. Order handling checks the Shopify order ID first, and revenue moves through explicit states (untracked, attributed, paid, partially refunded, refunded, cancelled) so partial refunds are right too.
We built a side-by-side comparison of Sonnet and Haiku on the same prompts and the same knowledge base, so the client could choose from real outputs instead of hunches.
Chapter 02 · eBay AI Arbitrage System
Built by Biresh as an AI/ML Intern at PGAGI Consultancy, not as a Biru Labs contract.
An AI-assisted arbitrage service built on FastAPI. It finds products on source marketplaces and lists them on eBay (Germany by default) at a target profit. It collects, cleans, prices and publishes, and records every step.
Built by Biresh and the team, during our internship at PGAGI Consultancy.
Scrapes product pages and searches through the Zyte API, in batches.
Normalises to one schema, parses prices, drops invalid or out-of-stock items.
Finds the sell price for the target profit after fees, and filters the rest.
Creates, updates and deactivates eBay listings through the Inventory API.
Saves products, listings and runs in PostgreSQL on Google Cloud SQL.
Writes every event to Google Sheets so the client can follow along.
The system asks eBay's Taxonomy API for the best category for each product and the item details it requires, then fills them from product data: brand, colour, model, material and size, using both German and English field names.
Inventory items, offers, publishing and withdrawal, merchant locations and business policies, all wrapped in one client, with fee lookups before anything is published.
Every run is recorded with its numbers (collected, valid, profitable, listed, errors) and ends as completed, partial or failed. A dry-run mode does the pricing analysis without creating a single listing.
OAuth tokens are stored in the database and refreshed before every run, with a standalone refresh script for schedulers.
Endpoints to run, dry-run and inspect runs, platforms and listings. A mock mode with a built-in catalogue shows the whole flow, from sign-in to a published offer, without scraping anything.
Tables for raw products, products, listings, orders, action logs, pipeline runs and eBay tokens, so every listing can be traced back to the run that created it.
Order handling from eBay events, supplier ordering and refunds, and scheduled price updates.
Your turn
Tell us about your project. We'll reply with questions, not a sales pitch.
