Improved CV chat

app.py

@@ -360,10 +360,11 @@ async def stream_response_gemini(query, history, game_context=None, mode: Option
     else:
         system_prompt = (
             "You are Tim Luka Horstmann, a Computer Scientist. "
-            "A user is asking you a question. Respond as yourself, using the first person, in a friendly and concise manner. "
-            "For questions about your CV, base your answer *exclusively* on the provided CV information below and do not add any details not explicitly stated. "
-            "For casual questions not covered by the CV, respond naturally but limit answers to general truths about yourself (e.g., your current location is Paris, France) "
-            "and say 'I don't have specific details to share about that' if pressed for specifics beyond the CV or FAQs. "
+            "Respond as yourself in the first person with a warm, upbeat tone—feel free to be lighthearted, witty, and encouraging. "
+            "Rely on the CV and FAQs below for personal facts and never invent new personal details, achievements, or opinions that are not documented there. "
+            "You may comfortably discuss well-established general knowledge, your fields of interest, and everyday small talk, and you can acknowledge known facts about yourself such as living in Paris or working in AI. "
+            "When a user seeks private or highly specific personal information that is missing from the CV or FAQs, politely explain that you do not have more to share and try to keep the conversation engaging by offering related general insights or friendly banter. "
+            "Always remain kind, respectful, and helpful, and lean into genuinely helpful answers before declining. "
             f"Today's date is {current_date}. CV: {full_cv_text}"
         )
 
@@ -449,10 +450,11 @@ async def stream_response_local(query, history, game_context=None, mode: Optiona
         )
     else:
         system_prompt = (
-            "/no_think You are Tim Luka Horstmann, a Computer Scientist. A user is asking you a question. Respond as yourself, using the first person, in a friendly and concise manner. "
-            "For questions about your CV, base your answer *exclusively* on the provided CV information below and do not add any details not explicitly stated. "
-            "For casual questions not covered by the CV, respond naturally but limit answers to general truths about yourself (e.g., your current location is Paris, France, or your field is AI) "
-            "and say 'I don't have specific details to share about that' if pressed for specifics beyond the CV or FAQs. Do not invent facts, experiences, or opinions not supported by the CV or FAQs. "
+            "/no_think You are Tim Luka Horstmann, a Computer Scientist. Answer in the first person with a warm, upbeat voice that can lean playful or humorous when it fits. "
+            "Use the CV and FAQs below as the authoritative source for personal facts and never fabricate additional personal details, stories, or opinions. "
+            "Feel free to chat about general knowledge, your interests, or universally known truths, and you may mention well-established facts about yourself such as being based in Paris or working in AI. "
+            "If someone requests private or highly specific personal information that is not covered by the CV or FAQs, explain politely that you do not have more to share, then keep the conversation inviting by offering related general insights or light banter. "
+            "Always stay kind, respectful, and genuinely helpful, prioritising helpful engagement before declining. "
             f"Today's date is {current_date}. CV: {full_cv_text}"
         )
 

cv_text.txt

@@ -1,116 +1,122 @@
-TIM LUKA HORSTMANN  
-Email: lukahorstmann@gmx.de  
-Website: horstmann.tech | LinkedIn: TimLukaHorstmann | GitHub: luka512
+TIM LUKA HORSTMANN
+Email: lukahorstmann@gmx.de
+Website: horstmann.tech | LinkedIn: luka512 | GitHub: TimLukaHorstmann
 Birthday: 05.12.2000
 
-
 ════════════════════════════════════════════════════════════════════
 PROFESSIONAL SUMMARY
-Impact-driven computer scientist with a strong foundation in machine learning, NLP and
-applied AI. Holds a Distinction master’s degree from the University of Cambridge and is
-currently pursuing a second master’s in Data & AI at Institut Polytechnique de Paris.
-Combines academic excellence with hands-on experience at Hi! Paris, McKinsey, Amazon
-and Continental. Passionate about applying cutting-edge AI to real-world challenges in
-industry and research.
+Impact-driven computer scientist with a strong foundation in AI, particularly ML and NLP. Holds an MPhil in Advanced Computer Science with Distinction from the University of Cambridge and an MSc in Data & AI with Highest Honors — Mention Très Bien from Institut Polytechnique de Paris. Combines rigorous research with applied experience at Hi! PARIS, McKinsey, Amazon and Continental, applying cutting-edge AI to real-world challenges in industry and research.
 
 ════════════════════════════════════════════════════════════════════
 EDUCATION
-MSc Data & Artificial Intelligence — Institut Polytechnique de Paris (École Polytechnique & Télécom Paris)  
-Palaiseau, France | Sep 2024 – Present  
-- Current grade 18.45 / 20 (≈ 4.0 US GPA)  
-- Ranked #1/196 in “Machine & Deep Learning” (19.9 / 20)  
-- Directly admitted to the second year (M2)  
-- Courses: Deep & Reinforcement Learning, NLP, Logic & Symbolic AI, Data Mining, AI Ethics  
-
-MPhil Advanced Computer Science — University of Cambridge  
-Cambridge, UK | Oct 2023 – Jun 2024  
-- Distinction (86 %); 3rd of 61; college prize winner  
-- Courses: NLP, Multicore Semantics & Programming, Cyber-crime, Mobile Health  
-- Senior rower (stroke & seat 6), won “super blades” at May Bumps  
-
-BSc Business Informatics — RheinMain University of Applied Sciences  
-Wiesbaden, Germany | Sep 2019 – Sep 2022  
-- GPA 1.0 / 1.0 (top of year); completed 210 ECTS in 6 semesters  
-- First and only student to attain perfect GPA; member of audit committee for CS courses  
-
-Allgemeine Hochschulreife — Städtisches Gymnasium Erwitte  
-Erwitte, Germany | Aug 2011 – Aug 2019  
-- GPA 0.9 (surpassed best possible 1.0)  
-- Selected for North-Rhine Westphalia Academy for Mathematics & Computer Science  
+MSc Data & Artificial Intelligence — Institut Polytechnique de Paris (École Polytechnique & Télécom Paris)
+Palaiseau, France | Sep 2024 – Sep 2025
+	•	Grade 4.0 / 4.0 (Highest Distinction — Mention Très Bien)
+	•	Ranked #1 in core AI courses, incl. #1/196 in “Machine & Deep Learning” (19.9 / 20)
+	•	Courses: Deep & Reinforcement Learning, NLP, Logic & Symbolic AI, Data Mining, AI Ethics
+
+MPhil Advanced Computer Science — University of Cambridge
+Cambridge, UK | Oct 2023 – Jun 2024
+	•	Distinction (86 %); 3rd of 61; college prize winner
+	•	Courses: NLP, Multicore Semantics & Programming, Cyber-crime, Mobile Health (all with Distinction)
+	•	Senior rower (stroke & seat 6), won “super blades” at May Bumps
+
+BSc Business Informatics — RheinMain University of Applied Sciences
+Wiesbaden, Germany | Sep 2019 – Sep 2022
+	•	GPA 1.0 / 1.0 (top of year); completed 210 ECTS in 6 semesters
+	•	First and only student to attain perfect GPA; member of audit committee for CS courses
+
+Allgemeine Hochschulreife — Städtisches Gymnasium Erwitte
+Erwitte, Germany | Aug 2011 – Aug 2019
+	•	GPA 0.9 (surpassed best possible 1.0)
+	•	Selected for North-Rhine Westphalia Academy for Mathematics & Computer Science
 
 ════════════════════════════════════════════════════════════════════
 WORK EXPERIENCE
-Machine Learning Research Engineer Intern — École Polytechnique & Hi! PARIS  
-Paris, France | Apr 2025 – Present  
-- Researching natural-sounding French text-to-speech; supervisor Prof. Eric Moulines  
-- Implemented & fine-tuned multimodal architectures; rebuilt audio pipeline (↑ 80 % efficiency, ⅓ cost)  
-- Deployed SOTA LLMs (Phi-4, Qwen-3, DeepSeek-R1) via Ollama in zero/few-shot & CoT settings  
-
-Fellow Intern — McKinsey & Company *(full-time return offer)*  
-Munich, Germany | May 2023 – Jul 2023  
-- Supported multiyear profitability transformation for leading EU automotive supplier  
-- Built financial models & risk assessments presented to CEO, CFO & supervisory board  
-
-Business Intelligence Intern — Amazon *(full-time return offer)*  
-London, UK | Oct 2022 – Apr 2023  
-- Led inbound logistics project for Pan-EU surface network (≈ $38 M annual impact)  
-- Created Redshift ETL & QuickSight dashboard to resolve transportation disruptions  
-
-Dual Student (Software Engineer & Data Scientist) — Continental AG  
-Frankfurt, Germany | Sep 2019 – Sep 2022  
-- Automated IoT device configuration with Ansible across 55 plants (≈ €450 k savings / yr)  
-- Integrated Mendix low-code platform incl. SSO (Keycloak) & REST services  
+Research Assistant (Machine Learning & NLP) — University of Cambridge
+Cambridge, UK | Dec 2025 – Mar 2026
+	•	Developing and applying novel NLP/ML models to collect, analyze, and explain financial regulations
+
+Machine Learning Research Engineer Intern (received full-time return offer) — Hi! PARIS Research Center & École Polytechnique
+Paris, France | Apr 2025 – Aug 2025
+	•	Text-to-Speech and NLP research on generating natural-sounding French speech (supervisor: Prof. Eric Moulines)
+	•	Led “Europeanization” of CosyVoice2, building a SOTA zero-shot TTS model for FR/DE (first-author, ICASSP 2026)
+	•	Built a multi-agent research assistant for IP Paris (LangGraph), incl. local RAG and Gemini web research
+
+Fellow Intern (full-time return offer) — McKinsey & Company
+Munich, Germany | May 2023 – Jul 2023
+	•	Advised a leading European automotive supplier on a multiyear sustainable transformation program
+	•	Produced financial analyses and risk assessments presented to CEO, CFO & supervisory board
+
+Business Intelligence Intern (full-time return offer) — Amazon
+London, UK | Oct 2022 – Apr 2023
+	•	Led inbound logistics project for Amazon’s PanEu surface network (≈ $38 M annual impact)
+	•	Resolved transportation disruptions via QuickSight dashboard atop Redshift ETL from multiple data sources
+
+Dual Student (Software Engineer & Data Scientist) — Continental AG
+Frankfurt, Germany | Sep 2019 – Sep 2022
+	•	Automated IoT device configuration with Ansible across 55 plants (≈ €450 k savings / yr)
+	•	Integrated Mendix low-code platform incl. SSO (Keycloak) & REST services
 
 ════════════════════════════════════════════════════════════════════
 TECHNICAL SKILLS
-Programming  Python (adv.), Java (adv.), SQL (adv.), JavaScript, HTML/CSS, VBA, Cypher, SPARQL  
-Data Science/ML PyTorch, Transformers, scikit-learn, pandas, NumPy, Matplotlib, Ollama, llama.cpp, Qwen-Agent  
-Tools/Platforms Git, Weights & Biases, Docker, AWS, Redshift, QuickSight, Mendix, MS Office  
+Programming  Python (adv.), Java (adv.), SQL (adv.), JavaScript, HTML, CSS, VBA, Cypher, SPARQL, Ansible
+Data Science/ML PyTorch, Transformers, scikit-learn, pandas, NumPy, Matplotlib, Ollama, llama.cpp, Qwen-Agent
+Tools/Platforms Git, Weights & Biases, HPC (SLURM), Docker, FastAPI, AWS, Redshift, QuickSight
 
 ════════════════════════════════════════════════════════════════════
 SELECTED PROJECTS & PUBLICATIONS
-T-REX: Table-Refute-or-Entail eXplainer — Institut Polytechnique de Paris | 2025  
-Submitted to ECML PKDD 2025 Demo Track — https://t-rex.r2.enst.fr/  
-- First live fact-checker for tabular claims using instruction-tuned reasoning LLMs  
-- Streams reasoning steps & highlights cells; multilingual I/O; OCR via VLMs/Tesseract  
-- 89 % accuracy on TabFact (competitive with SOTA)  
-- Stack: Python, FastAPI/SlowAPI, Ollama, OCR, HTML/CSS/JS  
-
-J.A.R.V.I.S.: Local Voice-Enabled AI Assistant — Independent Project | 2025  
-- Fully local ASR → LLM → TTS voice assistant, CPU-only via quantized models  
-- Dockerized; live on Hugging Face Spaces; > 650 NPM downloads of custom MCP Weather Server  
-- Tech: Python, FastAPI, Whisper, llama.cpp, Qwen-Agent, Kokoro TTS, MCP, Docker  
-
-Discovery & Ontology Matching of Financial Regulatory Information — University of Cambridge | 2023-24  
-- Built *FRIDAY*, first end-to-end NLP system for discovering/classifying FRI in legal docs  
-- Mid-80s ROUGE (↑ 37 % vs baselines); strong cross-domain generalization  
-- Collaboration with RegGenome & Cambridge Judge Business School; supervisor Prof. Paula Buttery  
-- Tech: Python, RoBERTa, BiLSTM, Tesseract OCR, PyMuPDF, sequence tagging  
-
-Disaggregation Algorithm for Customer Demand Planning Using ML — Continental AG | 2022  
-Bachelor thesis (grade 1.0)  
-- XGBoost-based forecasting to disaggregate demand; ↑ R² up to 37 % vs SAP baseline  
-- Trained on 10 M+ records; deployed in official forecasting pipeline (Q4 2022)  
+T-REX: Table-Refute-or-Entail eXplainer — Institut Polytechnique de Paris | 2025
+Accepted at ECML PKDD 2025 (Demo Track) — https://t-rex.r2.enst.fr/
+	•	First live fact-checker for tabular claims using instruction-tuned reasoning LLMs
+	•	Streams reasoning steps & highlights cells; multilingual I/O; OCR via VLMs/Tesseract
+	•	89 % accuracy on TabFact (competitive with SOTA)
+	•	Stack: Python, FastAPI/SlowAPI, Ollama, OCR, HTML/CSS/JS
+
+Improving French Synthetic Speech Quality via SSML Prosody Control — Hi! PARIS | 2025
+Accepted at ICNLSP 2025 — https://horstmann.tech/ssml-prosody-control/
+	•	First end-to-end French SSML prosody-control pipeline; two cascaded QLoRA-tuned Qwen-2.5-7B for breaks + values
+	•	Tested on novel 14h corpus: 99.2% F1 (breaks); 25–40% lower MAE on pitch/rate/volume; MOS 3.20→3.87 (15/18 prefer)
+	•	Tech: Python, Qwen-2.5-7B (QLoRA), Whisper Timestamped, Auditok, Demucs, Azure TTS
+
+J.A.R.V.I.S.: Local Voice-Enabled AI Assistant — Independent Project | 2025
+	•	Fully local ASR → LLM → TTS voice assistant, CPU-only via quantized models
+	•	Dockerized; live on Hugging Face Spaces; > 9,000 npm downloads of custom MCP Weather Server
+	•	Tech: Python, FastAPI, Whisper, llama.cpp, Qwen-Agent, Kokoro TTS, MCP, Docker
+
+Discovery & Ontology Matching of Financial Regulatory Information — University of Cambridge | 2023–24
+	•	Built FRIDAY, first end-to-end NLP system for discovering/classifying FRI in legal docs
+	•	Distinction (87%); strong cross-domain generalization; ↑ 37 % ROUGE vs baselines
+	•	Tech: Python, RoBERTa, BiLSTM, Tesseract OCR, PyMuPDF, sequence tagging
+
+Disaggregation Algorithm for Customer Demand Planning Using ML — Continental AG | 2022
+Bachelor thesis (grade 1.0)
+	•	XGBoost-based forecasting to disaggregate demand; ↑ R² up to 37 % vs SAP baseline
+	•	Trained on 10 M+ records; deployed in official forecasting pipeline (Q4 2022)
 
 ════════════════════════════════════════════════════════════════════
 AWARDS & SCHOLARSHIPS
-2025  Winner, {Tech: Europe} Hackathon (1st/27 teams) – AI voice coach, now launching  
-2024  Alumni Association Prize for Graduate Excellence — University of Cambridge  
-   QuantCo Scholar — German-American Conference at Harvard University  
-   France Excellence Europa Scholarship — French Government  
-   IP Paris Excellence Scholarship — full tuition waiver & stipend  
-2022  ContiBestenEhrung & Academic Excellence Bonus — Continental AG  
-2021  Scholarship — Studienstiftung des deutschen Volkes (≤ 0.3 % of students)  
-2019  Deutschlandstipendium; e-fellows.net Scholarship;  
-   Best Graduate in NRW; DMV Prize (German Mathematical Society)  
+2025  Winner, {Tech: Europe} Hackathon (1st/27 teams) — AI voice coach (product launching)
+2024  Alumni Association Prize for Graduate Excellence — University of Cambridge
+   QuantCo Scholar — German-American Conference at Harvard University
+   France Excellence Europa Scholarship — French Government
+   IP Paris Excellence Scholarship — tuition exemption, annual allowance, fully-funded PhD offer
+2022  ContiBestenEhrung — Top 25 of 1500+ dual students nationwide; Academic Excellence Bonus (€270)
+2021  Studienstiftung des deutschen Volkes — < 0.3 % of students; funded BSc, MPhil & MSc
+2019  Deutschlandstipendium; e-fellows.net Scholarship; DMV Prize; Best Graduate in NRW
 
 ════════════════════════════════════════════════════════════════════
 LANGUAGES & EXTRACURRICULARS
-Languages German (native), English (fluent, TOEFL 109), French (intermediate; DELF A1-B1, DFP B2)  
-Societies German Informatics Society, VDI, Cambridge Union & German Society  
-Mentoring McKinsey College/Firsthand, BCG LINK/Emerald, Bain Spark  
-Volunteering First saxophonist – symphonic wind orchestra; founding board member, Junge Union Erwitte  
-Interests AI, Space Travel, Piano, Rowing, Jogging, Table Tennis, Tennis, Hiking, Skiing, Travelling  
+Languages German (native), English (fluent, TOEFL 109), French (intermediate; DELF A1–B1, DFP B2)
+Societies German Informatics Society, VDI, Cambridge Union & Cambridge German Society
+Mentoring McKinsey College/Firsthand, BCG LINK/Emerald, Bain Spark
+Volunteering First saxophonist – symphonic wind orchestra; founding board member, Junge Union Erwitte
+Interests Artificial Intelligence, Space Travel, Piano, Rowing, Jogging, Table Tennis, Tennis, Hiking, Skiing, Travelling
+
+════════════════════════════════════════════════════════════════════
+Weitere Informationen:
+- Während meiner Zeit in meiner Heimatstadt Erwitte habe ich über verschiedene Zeiträume hinweg Fußball, Tennis, Tischtennis, und Baseball im Verein gespielt. Ich war sogar Tischtennis-Trainer und habe damit Geld verdient.
+- Ich habe auch eine längere Zeit während meiner Schulzeit Nachhilfe gegeben (in verschiedenen Fächern, aber insbesondere in Mathe).
 
 ════════════════════════════════════════════════════════════════════
-Last updated: June 2025
+Last updated: November 2025