Prince Sinha

Backend engineer. Shipping multiple B2B SaaS products solo.
Writing about production AI, billing infra & complexities of building software. Open to remote work

[Prince Sinha @just_figuring]

@walden_yan The VM-with-apps-preinstalled approach vs. blank sandboxes is the unlock most teams haven't internalized yet. Giving the model a computer that's already "onboarded" to the repo and toolchain saves more tokens than any prompt engineering ever will.

[Prince Sinha @just_figuring]

@DanWahlin The breadboarding step is what most people skip and then wonder why autopilot goes off the rails. Sketching the flow before defining tasks forces you to find the gaps in your own thinking before the agent has to.

[Prince Sinha @just_figuring]

@swyx @ACM_President @CAISconf @aiDotEngineer "Turing of AI Engineering" should reward shipped systems, not papers. The thing that's missing in this field is recognition for the people who figured out how to make eval pipelines, retrieval, and agent orchestration actually work in prod not just the ones who published about it

[Prince Sinha @just_figuring]

@coltpdx Played it. Smooth 60fps on mobile Safari with 20+ pieces in motion is not trivial , most browser games tank way before that. Curious if you're using Canvas2D or WebGL for the rendering.

[Prince Sinha @just_figuring]

@simonw Genuinely useful eval. The fact that you can see thinking effort matter visually, not just on a benchmark number, is the best argument for these tiers existing.

[Prince Sinha @just_figuring]

@levelsio The George Hotz thesis aged like fine wine. Most people laughed at that take in 2024.

[Prince Sinha @just_figuring]

@jamesqquick Nice. Does it support percentage rollouts and targeting rules (per user/region), or is it boolean-only for now? That's usually what pushes me to third-party tools.

[Prince Sinha @just_figuring]

Mental model for any retry PR:

"What if 1000 of these fire at the exact same moment?"

If the answer isn't "nothing bad," it needs jitter, a breaker, or a budget. Usually all three.

Exponential backoff without jitter isn't backoff. It's a synchronized DDoS on a timer.

[Prince Sinha @just_figuring]

One of the most common traps in distributed systems looks innocent in code review:

time.sleep(2 ** attempt)

"Exponential backoff." Textbook. Ships in thousands of services.

Then one bad day, it takes the DB down. Here's the mechanic. 🧵

[Prince Sinha @just_figuring]

The reason this bug is so common isn't carelessness — staging hides it.

Synchronization is invisible at small N. With 4 workers, sleeps drift. With 600, they line up.

The failure mode doesn't exist until concurrency × error-rate crosses a threshold.

[Prince Sinha @just_figuring]

Spent 6 hrs building a custom RAG pipeline. Fixed by switching to Claude's native document context. Now I just pass the PDF directly. No embeddings, no vector DB, no infra. Works for 90% of real client needs.

[Prince Sinha @just_figuring]

Unpopular: MCP servers are overkill for 80% of agent use cases. Tested on 3 client projects → plain function calling + a well-structured prompt did the same job. Less config, fewer moving parts, ships faster.

[Prince Sinha @just_figuring]

Indian devs are charging $25/hr for work that US freelancers bill at $120. Same stack. Same output quality. The gap isn't skill — it's positioning and proof. Fix the portfolio before you fix the rate.

[Prince Sinha @just_figuring]

Claude Code just rewrote a 200-line Express route I wrote. Kept the logic, killed the boilerplate. Diff was -87 lines. The uncomfortable part: I couldn't find anything wrong with it. #buildinpublic

[Prince Sinha @just_figuring]

Source 3: Previous clients. Just stayed in touch — shared something useful once a month. No pitch. Both repeat engagements came inbound. Referrals beat cold outreach every time. Build the relationship, not the funnel.

[Prince Sinha @just_figuring]

I track every hour I bill. Last 90 days as an Indian freelancer working global clients: $0 from job boards. 100% from 3 sources. Here's what actually worked → 🧵

[Prince Sinha @just_figuring]

Source 2: Cold emails to bootstrapped SaaS founders (not agencies). Researched their stack first. Proposed one specific fix. Reply rate: ~14%. Conversion: 2 clients from 30 emails. Low volume, high intent.