Computer science, UC Irvine. Co-founder, Crestfora.
I’m Theo. I study computer science at UC Irvine and run Crestfora, a B2B appointment-setting agency, with my co-founder Mark Phouiphiphath. I also build Python tools for machine learning and language models. Every project below links straight to its code.
Crestfora helps B2B companies get meetings with the decision-makers they want to reach. We research the right contacts, start the conversation with personalized outbound email, qualify the replies, and book sales meetings onto our clients’ calendars.
Co-founded HVACNow, an AI automation business focused on helping HVAC companies recover missed revenue.
Built the initial version of our AI-powered phone system, including bilingual English/Spanish call handling, emergency escalation rules, lead qualification, and CRM integrations. Worked alongside my co-founder to develop the product, speak directly with business owners, and refine the offer based on what we learned.
Led the early sales and customer discovery process, from prospect research and cold calling to testing our positioning and identifying where automation could provide real value.
HVACNow was founded in 2026 in Irvine, California by Theo Markopoulos and Mark Phouiphiphath. The founders later wound the company down and pivoted into Crestfora.
Trains lightweight classifiers on a language model’s internal activations to flag likely hallucinations while the model is still generating, with no retrieval step. It implements the method from a 2026 research paper and trains on SQuAD v2 with weak-supervision labels.
75handle = layer_module.register_forward_hook(_hook)
Watches a stock watchlist and scores signals from live market data, technical indicators, machine-learning anomaly detection and news sentiment, then texts an alert when a signal is strong. It recommends; it doesn’t trade.
45raw_kelly = (b * p - q) / b
A routing layer for multi-agent LLM pipelines. It classifies each sub-task and uses a bandit algorithm to pick the cheapest model likely to handle it, logging every decision so the choices improve over time.
126return max(affordable, key=lambda k: self._ucb_score(k)).model_id
A FastAPI service that scores how reliable a reasoning model’s answer is from its reasoning trace alone. It counts hedging and verification phrases to compute a Hedge-to-Verify Ratio and a confidence score, then returns accept or defer. It can also sit in front of any OpenAI-compatible API as a proxy. It implements the scoring method from a 2026 research paper.
19hvr = h / (v + 1)
An evaluation suite that measures how often language models refuse legitimate requests involving rules that are defeated, unjust or absurd. It builds a 95-scenario test matrix across 5 defeat families and 19 authority types, runs it through LiteLLM, and scores both over-refusal and whether the model recognizes why the rule doesn’t apply. It implements the method from a 2026 research paper.
604for family, authority in itertools.product(families, authorities):
Extracts a capability, such as chain-of-thought reasoning, as a direction in one language model’s residual stream and transfers it into another model in the same family through linear alignment, with no fine-tuning. It implements the method from a 2026 research paper.
113w, *_ = np.linalg.lstsq(src_acts, tgt_acts, rcond=None)
B.S. in Computer Science. Coursework in data structures, algorithms, systems programming and MIPS assembly.
Led shifts of up to 10 people and trained new hires. Built an Excel labor report that saved about three hours of reporting a week, and introduced a handoff check that cut order errors by about 20%.
Built ZotLabs, a research-matching platform for UCI students and researchers, in a nine-person team. I implemented role-based routing and core UI components, and debugged auth and session edge cases. React, Vite, Supabase and Tailwind.
Started on a pre-med track, switched to computer science, and transferred to UC Irvine.
For my Eagle project I planned and led eight volunteers in building a shade structure at a nonprofit animal shelter.