ML + AI Agent Engineer · Ex-Meta Ads ML · Ex-Apple Face ID · CTO @ Aonxi Santa Monica, CA · origin@aonxi.com · aonxi.com · linkedin.com/in/samisanmol

AGI agents are not chatbots. They are not copilots.

They are revenue-making, P&L-responsible autonomous systems that compound value the way a senior operator would - but at infinite scale and zero marginal cost.

Imagine giving your agent an unlimited corporate credit card. It can buy any API, spin up GPUs, run experiments, hire compute. But it must bring qualified meetings, pipeline, and closed-won revenue. Every dollar it spends must return more dollars than it cost.

That is not a demo. That agent exists. It is running right now. It generated $650K ARR in 5 months. Bootstrapped. Zero paid acquisition.

I am the person who built it. And I am just getting started.

Most people talk about AGI in the abstract. I am building it in the concrete - one bounded domain at a time.

The domain I chose is revenue and capital. Not because it is easy. Because it has the clearest eval in existence:

Did the meeting happen? Did the deal close? Did money land in the bank?

No ambiguity. No benchmark gaming. No cherry-picked demos. The production system is the laboratory. The ARR is the p-value.

Stage 1 (Live now): Lead outreach and discovery The agent wakes up every morning. It searches 30+ data sources for companies matching the ICP. It researches each one - funding news, hiring signals, product launches, founder posts, acquisition events. It writes a hyper-personalized email proving it did the research. It scores intent using dual-LLM judgment (Claude + Grok, weighted average). Human reviews. Human approves. Email sends. Meetings land on the calendar.

Stage 2 (Building now): Intelligent follow-up and sequencing Three-touch sequences. Fresh signals on each follow-up. Reply detection via IMAP. Auto-routing of warm responses. The agent learns which angles convert by vertical, size, signal combination. Every outcome becomes an anchor in the memory bank.

Stage 3 (Next): Shadow sales and learning The agent observes real sales calls. It extracts objections, winning tactics, buyer psychology, pricing signals. It builds a model of what closes - not from theory, from real conversations. It feeds this back into outreach to close the loop between lead and close.

Stage 4 (Coming): Sales assistance and co-piloting The agent co-pilots sales reps in real time. Surfaces the right objection handler at the right moment. Drafts follow-up emails while the call is still happening. Escalates only when the deal size or complexity demands human judgment.

Stage 5 (The destination): Full autonomy First touch to contract signed to money in the bank. The agent runs the entire revenue funnel. It spends more where results compound. Kills what does not work. Requests human input only when it creates leverage. Every decision is logged, attributed, and auditable.

This is AGI for a bounded domain. I am building it step by step, in production, with real money on the line.

I am not a researcher who builds demos. I am not a founder who hires engineers. I am both - and I can connect everything to everything.

At Meta I shipped ranking and bidding models at billions of daily inferences. At Apple I reduced Face ID false positives by 15% and published the GAN refiner that made synthetic training viable (machinelearning.apple.com/research/gan). At Lotus Interworks I took ARR from $25K to $4M in 9 months, cut the sales cycle from weeks to under 5 days, and lifted close rate from 10% to 34%. At Aonxi I built the agent from scratch. Alone. $0 raised. $650K ARR in 5 months.

The combination nobody else runs at scale: Deep ML research + agent systems engineering + production revenue data

I replicate frontier papers the week they drop. I deploy the architecture into the live system. I measure whether the theory holds when money is on the line. If it does not, I fix it or throw it out.

This is not how most researchers work. This is not how most engineers work. This is how you build AGI that actually does something.

An agent with an unlimited credit card and no constraints is not AGI. It is a liability.

The architecture I engineered to keep it safe, auditable, and production-grade:

Multi-agent swarms with 4-layer quality gates No single agent makes a final decision. Every action passes through intent verification, confidence scoring, compliance checking, and human escalation logic before execution.

Human-in-the-loop (HITL) nodes everywhere Y/N vetoes. Approval gates. Override controls. Rich feedback loops. Every human decision becomes a training signal. Low-confidence cases escalate to humans automatically. Human labels feed continuous iteration. This is RLHF running live in production, every day.

Uncertainty-based routing When the agent is not sure, it asks. When it is sure and wrong, it learns. The confidence threshold for autonomous action rises only when production eval data supports it.

Economic ledger with hard gates Every execution logs cost per task, variance, and outcome delta. Scaling is gated on probabilistic ROI - not task success percentage. No logs, no scale. No proven ROI, no autonomy increase.

Kill switches at every layer Agents are stateless and short-lived. They can be terminated mid-task without corruption. The orchestrator owns workflow state - not the agents. This means I can pull the plug on anything, instantly, at any layer.

Winners auto-replicate and compound. Losers terminate with zero drama. Better data beats better algorithms. I have lived this truth at Meta scale.

Every paper I read gets deployed. Every architecture I replicate gets tested on production revenue data. Here is what I am working from and why each one matters for AGI:

AndroTMem (arxiv:2603.18429, March 2026)

The problem it solves: Long-horizon agents fail primarily because of memory failures. Not intelligence. Memory. As step count grows, agents forget causally important intermediate states. They repeat November's mistakes in January. They cannot learn from December's success.

The December anomaly - real production data:

Same agent. Same verticals. 10.5x efficiency difference. The agent had no idea why December worked. That observation became this research program.

My contribution over the paper: The original handles single-session tasks (max 65 steps). I handle multi-session campaigns (6 months, 2,452 leads, 42 sales). First production validation of ASM on real financial outcomes.

Real experiment results:

ASM beats raw scoring on 83% of prospects. The December "quality over quantity" insight lives as a causal anchor. The agent retrieves it when scoring similar prospects. It will never repeat November again.

→ github.com/originaonxi/asm-replication

Why it matters for AGI: Outcome reward is not enough. An agent that only learns from final results cannot improve the quality of its reasoning process. PRM rewards intermediate steps - it teaches the agent to think well, not just to get lucky. For the revenue agent: did the scoring decision at step 3 contribute to the meeting at step 47? That attribution chain is what PRM formalizes.

→ github.com/originaonxi/prm-replication

Why it matters for AGI: Single-step reward is local. RewardFlow propagates reward signals across the full action chain - crediting or penalizing decisions that caused outcomes several steps later. Revenue attribution IS RewardFlow in the wild. Every email sent, every follow-up skipped, every angle chosen - each one contributed to whether the meeting happened. RewardFlow is how you learn which contribution mattered most.

→ github.com/originaonxi/rewardflow-replication

Why it matters for AGI: Trajectory data is the ground truth of agent behavior. Not what the agent said it would do - what it actually did, step by step, under real conditions. For the revenue agent: the full trajectory from "Apollo returned this company" to "meeting booked" is the dataset. TDAD is the framework for learning from it.

→ github.com/originaonxi/tdad-replication

Why it matters for AGI: State is the unsolved problem of production agents. What do you keep? What do you compress? What do you discard? How do you structure it so retrieval is fast, relevant, and causally grounded when the context window runs out? This thesis formalizes those decisions for agents operating over thousands of steps and months of campaigns.

→ github.com/originaonxi/information-architecture-thesis

Routing Strategy and Load Balancing in Small-Scale MoE Transformers

Controlled comparison of top-k, expert choice, and noisy top-k routing. Original expert specialization score metric. Key finding: Top-2 routing + light auxiliary loss delivers 20% perplexity improvement and prevents collapse. Directly applicable to multi-agent swarms where routing decisions determine which persona model handles which prospect.

→ github.com/originaonxi/moe-efficiency-study

The bottleneck is not intelligence. It is reliability.

GPT-4 can reason. Claude can reason. Gemini can reason. None of them can operate a live business autonomously for 6 months without accumulating failures that compound.

The five components you need to build AGI that actually does something:

Memory that survives sessions (ASM) Not context windows. Structured causal anchors. What happened, what caused it, what it means for next time. Reward that propagates through time (PRM + RewardFlow) Not "did this email get a reply" but "did this decision pattern produce meetings across 6 months of campaigns." Trajectories that stay coherent (TDAD) An AGI that drifts after 100 steps is not AGI. It is a very expensive random walk. Human-in-the-loop as a learning signal (HAGI) Not a safety patch. A continuous preference signal. Every Y/N approval is a labeled training example. That is RLHF running live in production, every day. Production data as ground truth (Aonxi $650K ARR) Benchmarks can be gamed. Revenue cannot. 2,452 leads. 250 meetings. 42 sales. Real money. That is the eval.

Put these five together and you get: An agent that improves itself continuously, operates reliably over months, attributes outcomes to decisions, and gets better every single day without human intervention.

That is AGI for a bounded domain. I chose revenue because the eval is unambiguous. I am building it in production because that is the only environment where the theory gets tested honestly.

I operate at the intersection where nobody else stands: Visionary enough to see the destination. Technical enough to build the path. Obsessive enough to ship it daily.

The names people invoke when they talk about this level of ambition - Elon, Demis, Sam, Andrej - they are all right about where this goes. The difference is I can also write the code, connect the agent to everything, and put real money on the line to prove it works.

Morning: Read one new paper or arxiv preprint Run aonxi-outreach (20+ new emails sent) Note what the agent got wrong Afternoon: Review replies, close clients Run aonxi-sequence (follow-up sequences) Log outcomes as ASM anchors Evening: Build or improve one thing Push to GitHub Connect what I built to a paper I read Weekly: Write up what the data showed Update paper outline with new numbers Ask: what would falsify this hypothesis? Always: Every paper connects to a live experiment Every experiment produces real outcome data Every outcome improves the next decision

Agent core: Python · Anthropic API (Claude Haiku/Sonnet) Grok API · dual-LLM weighted scoring Discovery: Apollo · Exa · Hunter · 30+ source waterfall Memory: SQLite ASM anchor bank · structured causal retrieval Delivery: Gmail SMTP · SendGrid · IMAP reply detection Orchestration: Temporal.io · durable execution · kill switches Infra: Kubernetes · spot GPUs · Rust + Tokio core Data: Kafka/Redpanda · ClickHouse · 100M+ B2B records Research: PyTorch · vLLM · Qdrant · LoRA fine-tuning Observability: Custom revenue attribution dashboard · OpenTelemetry HITL cockpit: React · WebSockets · Y/N/feedback nodes

Aonxi - CTO (Sep 2025 – Present · Santa Monica) Built the autonomous revenue agent from scratch. $650K ARR in 5 months. $0 raised. Zero paid acquisition. 120+ revenue-trained persona models as live autonomous workflows. 76 enterprise-grade real-time intent filters. 100M+ continuously enriched B2B records. Native HITL nodes: veto, approve, override, feedback. Every architecture decision tested on real production outcomes.

Lotus Interworks - VP/Director Revenue Engineering (2023–2025 · LA) ARR: $25K → $4M in 9 months. Close rate: 10% → 34%. Sales cycle: weeks → 5 days. Pipeline throughput: 10x. Qualified pipeline exceeded $100M. Built PilotOS (Results-as-a-Service) treating revenue as a constrained learning control system. Co-designed ECOSYSTEM framework with Dr. Gopinath (Stanford PhD) and Ray Stata (Founder, Analog Devices).

Meta - ML Engineer, Ads (2017–2023 · 6 years · Menlo Park) CTR/CVR prediction, ranking, and bidding at billions of daily inferences. Data augmentation, synthetic data, knowledge distillation. Early production PyTorch eager-mode adopter - shipped one of the first ad-ranking models on dynamic graphs when the industry was still on static TensorFlow.

Apple - Research Scientist ML (Intern) (2017 · Cupertino) Reduced Face ID false positives by 15% on edge cases. Built sim-to-real GAN refiner for synthetic training. Published: machinelearning.apple.com/research/gan (arXiv 1612.07828).

AutoScholar Pro - Founder (Age 10–20 · India) Self-taught on a second-hand PC. No internet. QBASIC books from the library. 12-hour daily sprints. Built an automated university and scholarship platform serving 5,000 students, unlocking $2M+ in scholarships. 90% acceptance prediction accuracy. Exited for $100K. First line of code written at age 10. Have not stopped shipping since.

Northwestern University - Kellogg School of Management, MBA (Marketing) Delhi University - B.Sc. Mathematics Frontier AI Open School - Andrej Karpathy directed Deep Learning & LLM Systems

I want to talk to people working on:

• Long-horizon agent reliability in production
• Memory architectures beyond context windows
• RLHF from live human-in-the-loop systems
• Production evals where real outcomes are the metric
• AGI that operates in bounded real-world domains
• Agents that are P&L accountable, not just task-complete

Frontier labs building AGI - I am your hire. I can see where this goes and I can build the path to get there. I have the production system, the research program, the engineering depth, and the evidence.

origin@aonxi.com · aonxi.com Santa Monica, CA · Shipping every day since age 10