Binocs - the company and your role

What Binocs actually is

Binocs is a commercial due diligence platform for private equity associates, corporate development teams, and consultants. The job-to-be-done is brutal in its old form. A deal lands on a Friday. The associate has the weekend to read a 200-page CIM, build a market map, score the moat, find three customer references, write a thesis memo, and put it in front of the partner Monday morning. The old way is six analysts and a lot of coffee. The Binocs way is a few prompts, a structured workflow, and an output the partner can actually read.

The product is not "ChatGPT for PE". The reason it works is the structured CDD workflow on top of the LLM, the verticalized templates per industry (SaaS CDD, fibre infra CDD, healthcare CDD), and the eval set built from real shipped deals. The LLM is the engine, the workflow is the car.

How I ended up there

I got in through a referral and a take-home. The take-home was a slide-deck parser. I shipped it in two days, then shipped a v2 nobody asked for that handled rotated text. That bought me the interview, and the interview was a system design (rate-limit a webhook fanout) and a debugging round on a real PR. They made the offer the same week. July 2025, in-person Bangalore.

The first project - Stripe

The day I joined, the Stripe integration was a single file and a prayer. International payments worked for US customers, mostly. Tax was hardcoded. Webhooks were unverified. The PM wanted to charge an EU customer by end of month.

I split the integration into three services - checkout, tax-and-geo pricing, webhooks. Idempotency keys everywhere. Signature verification on every webhook. A reconciliation job that diffed Stripe state against our DB every hour and flagged drift. End of year, zero failed transactions in prod. The deeper story is in the Stripe topic, but the headline is that the team trusted me with money on day 30, and that shaped every project after.

The deck pipeline - 99 percent cost cut

The slide deck pipeline turned a CIM PDF into a structured deck the associate could edit. The naive version sent every page to GPT-4 with "extract the structure and write the slide". At our volume that was burning thousands of dollars a month and producing inconsistent output.

I rewrote the pipeline so the LLM only did the parts the LLM was actually good at - the writing. Layout, image extraction, table parsing, bullet ordering, chart re-rendering - all deterministic code with a real parser, not an LLM. The model only got the cleaned, normalized text and a tight prompt. Cost dropped 99 percent. Latency dropped too, because deterministic code is faster than waiting on a model. The deep dive is in the deck pipeline topic.

EKS - the boring win

A senior asked me to "look at the AWS bill". I spent a week with Cost Explorer, CloudWatch metrics, and the EKS node groups. Findings, in order of impact - we were running m5.2xlarge nodes at 18 percent CPU, RDS was on a tier two sizes above what it needed, we had no spot instances for batch jobs, and one dev cluster had been left on for three months.

Right-sized the node groups, moved batch to spot with a fallback, dropped RDS one tier with a soak test, killed the orphan cluster. 1.8K to 2K USD per month saved, on a small team that is real money. Cumulative 60 percent on the relevant line items. Deep dive in the EKS topic.

The 24h admin tool

The internal ops team was using a Google Sheet plus three different Retool dashboards to manage user invites, deal access, and feature flags. The lead engineer asked me on a Tuesday evening "can we just have one admin panel". I said yes by Wednesday evening.

I shipped a Next.js admin app with NextAuth, role-based access control, full CRUD on users and deals, feature flag toggles, and an audit log. 40-plus internal users were on it by Thursday morning. The trick was not magic, it was scope discipline - I shipped exactly the screens they needed and nothing else. Deep dive in the admin tool topic.

Fibre CDD vertical - 7 to 8 days

The CEO wanted to pitch a fibre infrastructure deal in 10 days and we did not have a fibre vertical. I worked with the domain expert (an ex-fibre analyst) to write the templates, the eval set, and the workflow. 7-8 days from kickoff to demo. We won the pilot.

The lesson was that verticalization is not a model fine-tune, it is a workflow and an eval set. The same base model, with the right templates and the right grader, becomes a domain expert.

The culture

Two things made Binocs different from every internship I had heard about.

First, deep ownership. Nobody assigned me Stripe. The lead said "we need payments to work, what do you want to do" and I said "all of it" and that was the conversation. Same for the deck pipeline. Same for EKS. If you asked for surface area, you got it. If you did not ask, you got tickets.

Second, the pace. Weekend hot fixes were not a fire drill, they were Tuesday with a different calendar. The product was in front of paying customers, the customers were PE associates with billion-dollar deals, and a bug in prod was a real bug. You did not learn to ship by reading about it.

Why I left

I declined the FTE in May 2026. The work was the best I had done. The team was the best I had worked with. The reason was simple - my father has stage 3 cancer, treatment is ongoing, and Bangalore is not Delhi. I needed to be near family, full stop. I told the lead the truth, they understood, the offer stayed open in case my situation changed.

What I would tell the next intern

• Ship the first thing they hand you on day two, not day five. Speed buys trust, trust buys surface area.
• When you do not know how to solve a problem, do not say "I do not know", say "give me 30 minutes". Then come back with options.
• Read the prod logs. Every day. You will find bugs before customers do, and you will earn a reputation as the person who finds bugs before customers do.
• Take notes. Write down what broke and how you fixed it. Future you will need it.
• Push back on scope. The 24h admin tool worked because I cut features, not because I worked faster.

That is Binocs. The next nine topics are the projects in detail.