Skip to content
← Back to Overview / Live Command Deck
OVERFITTED SUIT-UP
Subscribe Free
STAGE 01 // AGENT 01

Stark Telemetry Dais & Titanium Endoskeleton

STAGE 1/8
INITIALIZING IRON SUIT BLUEPRINT...
SUIT-UP INTEGRATION0% ASSEMBLED
Munna Tactical Officer

MUNNA // SUIT-UP PROTOCOL

"Telemetry dais active. Internal titanium chassis locked onto Bengaluru grid."

THE BUILD // MARCH 2026

Built from Bengaluru. Assembled in public.

Scroll down to watch the Overfitted autonomous combat suit build itself piece-by-piece from nothing—like an Iron Man suit-up scene locking together in mid-air—while discovering the authentic struggles, post-mortems, and engineering triumphs of running a solo AI publication.

SCROLL DOWN TO INITIATE SUIT-UP ASSEMBLY
01 // STAGE 01 // AGENT 01

March 2026: The Noise Floor & The Blueprint

Surveillance Array

The signal was buried. So we built an intake firehose.

In early March 2026, in Bengaluru, Overfitted was born out of frustration. The AI space had turned into a breathless PR swamp: five hundred papers indexed every day, AI newsletters summarizing each other's hallucinations, and genuine breakthroughs buried under superficial hype.

Aseem asked a radical question: Can one person build a media operation that operates with the rigor of a research lab, runs 24/7 autonomously, but never lies?

The Stark assembly pedestal ignites. The internal titanium spinal chassis and rib framework lock into the glowing telemetry platform, establishing the foundational skeleton for all 8 autonomous checkpoints.

The Struggle // Where It Broke

The Initial Trap: Pulling 150+ raw articles an hour almost crushed our parser. We had to build custom pre-filters to throw away 90% of PR fluff before a single token was processed.

INGESTION15 Feeds / hr
CANDIDATES142 parsed / run
DISCARD RATE88% PR noise
02 // STAGE 02 // AGENT 02

Memory That Never Forgets

Novelty Gate

Preventing self-repetition without human memory.

The quickest way an autonomous newsletter dies is by repeating itself. LLMs have no persistent long-term memory; left alone, they will enthusiastically write about the same topic three weeks in a row.

The segmented abdominal armor plates slide over the waist, while twin high-velocity rear flight thruster pods dock onto the spine. Every story candidate is converted into a 1,536-dimensional dense embedding in Qdrant.

If the cosine similarity to any previously published issue is greater than 0.40, the story is killed on sight. No repeats. No echoes.

The Struggle // Where It Broke

Tuning the Distance Threshold: Set it too high (0.60), and readers get duplicate angles. Set it too strict (0.25), and follow-up breakthroughs get blocked. 0.38–0.40 became our gold standard.

VECTOR DBQdrant Cloud
ARCHIVE SIZE1,280 Vectors
THRESHOLDCosine > 0.40 Fail
03 // STAGE 03 // AGENT 03

The Dirty Reality of Web Scraping

Evidence Locker

Where things broke: Paywalls, SPAs, and truncated PDFs.

AI models hallucinate when their context is starved. If an agent only gets an article title and a 200-word snippet, it will invent the rest with cheerful confidence.

Heavy contoured shin greaves and armored thruster boots fly in from below, clamping onto the legs. Dedicated extraction workers powered by Trafilatura crawl into primary PDFs, GitHub READMEs, and technical docs to harvest full source text.

This is where production broke hardest early on: Cloudflare bot walls, bloated client-side SPAs rendering blank HTML, and arXiv PDFs failing to parse. We hardened the scrapers to extract verified primary text or fail the entire build.

The Struggle // Where It Broke

Post-Mortem: A broken scraper once passed an empty markdown string to the writer, resulting in an essay about 'Cookie Policy Accepted.' We instituted a fail-closed gate: minimum 800 tokens of verified body or instant abort.

EXTRACTORTrafilatura Engine
MIN BODY800 Verified Tokens
PRIMARY CITATIONS12–18 per story
04 // STAGE 04 // AGENT 04

The Micro-Budget Breakthrough: $23/Month

Cold Synthesis

Demolishing the myth that AI media requires VC millions.

Every VC-backed media startup burns tens of thousands of dollars on expensive enterprise clusters and cloud GPU farms. Overfitted proved the opposite: the entire stack runs autonomously on a single $23/month Hostinger VPS.

The massive angular pectoral chestplate slams shut over the ribs, and the Arc Reactor ignites with brilliant cyan/green light. Powered by DeepSeek-V3, our structured synthesis agent consumes the evidence brief under strict constraints: 'You cannot assert any fact that is not traceable to an Agent 03 citation token.'

Average token cost per newsletter issue: $0.023. Not twenty-three dollars—two point three cents. Proof that scale is no longer an economic moat.

The Engineering Win

Economic Victory: 3 newsletters a week, autonomous background daemons, zero cloud lock-in. A complete media company running for less than a family dinner in Indiranagar.

INFRASTRUCTUREHostinger VPS ($23/mo)
TOKEN COST$0.023 / issue
LLM ENGINEDeepSeek-V3
05 // STAGE 05 // AGENT 05

The Scariest Failure: The Sentence Guillotine

Hallucination Scan

When a model invented a benchmark that sounded completely real.

In late March, during an experimental run, the synthesis model generated an astonishing claim: that a new open-source model beat Claude 3.5 on a benchmark called 'SWE-Lattice.' The sentence was written with flawless prose, citations, and authority.

There was just one problem: 'SWE-Lattice' did not exist. The model had seamlessly hallucinated a plausible-sounding benchmark name.

Arm gauntlets sweep in with glowing palm repulsor lenses, while diagnostic amber laser beams sweep vertically down the armor. The Hallucination Guillotine slices the draft into individual sentences; any unreferenced assertion is quarantined immediately.

The Struggle // Where It Broke

Zero Tolerance: No human can read fast enough to catch microscopic factual hallucinations under deadline. The machine must act as its own harshest skeptic.

METHODPer-Sentence Slicing
BYPASS POLICYZero Bypass (Fail-Closed)
QUARANTINESub-assembly ejected
06 // STAGE 06 // AGENT 06

The Automated Second-Chance Engine

Fact Validation

Re-verifying flagged claims or purging them forever.

Quarantining a claim isn't enough; the system must know whether the claim was a false positive or a lethal hallucination. Heavy quadrilateral armor plates slide onto the upper thighs, fused with molten-gold Kintsugi joints.

Agent 06 dispatches targeted, independent web queries against secondary authority databases (arXiv tables, company GitHub commits, verified press).

If the claim can be independently anchored by a second authoritative source, the suture heals and the claim is reinstated. If not, it is permanently deleted from the final draft. Zero hand-waving.

The Engineering Win

Kintsugi Principle: Just like the Japanese art of repairing broken pottery with gold, our pipeline becomes stronger precisely where it was stressed.

VALIDATORAutomated Secondary Search
PASS RATE94.2% on flagged
PURGE RATE5.8% removed forever
07 // STAGE 07 // AGENT 07

Never Trust a Single Model's Ego

Consensus Check

Adversarial audit by an independent second LLM.

Even a factually accurate draft can suffer from subtle tonal bias, sloppy phrasing, or overconfident extrapolations. Massive curved shoulder pauldrons slam down onto the collar, giving the suit broad, imposing proportions.

An independent second-opinion LLM (adversarial reviewer) receives the assembled draft with zero knowledge of the generation prompt. Dual counter-rotating gyroscopic rings spin around the torso to verify 95% alignment.

If the consensus alignment score is below 95%, the draft is rejected back to the synthesis chamber. Only drafts with mathematical consensus pass through.

The Struggle // Where It Broke

The Model Bias Trap: Different LLMs have different stylistic blind spots. Using a different model family for Agent 07 ensures one vendor's quirks don't slip through unnoticed.

AUDITORIndependent Model B
THRESHOLD>= 95% Editorial Consensus
ALIGNMENT98.4% achieved
08 // STAGE 08 // AGENT 08

Mon / Wed / Fri 7:00 AM IST: The Send Button

Human Sign-Off

Machines handle the rigor; a human gives it soul.

The grand climax: The aerodynamic helmet descends from above onto the neck collar. The golden faceplate rotates down and magnetically SLAMS shut with an echoing metallic lock.

The piercing white-cyan visor eyes ignite! The palm repulsors pulse forward, boot thrusters blast golden particle exhaust, and the entire suit locks into full combat flight readiness.

At 6:30 AM IST, Aseem reviews the telemetry logs, inspects the verified proofs, and hits dispatch. At 7:00 AM IST, 1,500+ engineers, researchers, and founders receive pure, verified truth in their inbox.

The Engineering Win

Proof of Concept: 44+ issues shipped, 1,500+ subscribers, 0 hallucinated errors slipped to production. A one-person media company competing at world scale.

CADENCEMon/Wed/Fri 7:00 AM IST
READERS1,500+ Subscribers
MANUAL STEPExactly 1 (Hit Publish)

The Suit is Primed. The Proof is in the Inbox.

Published every Monday, Wednesday, and Friday at 7:00 AM IST. No hype, zero hallucinations, strictly verified.

Subscribe to Overfitted FreeReturn to Home